Discussion Decision Making Week 4

RESEARCH ARTICLE

Dual Processing Model for Medical Decision- Making: An Extension to Diagnostic Testing Athanasios Tsalatsanis1,2, Iztok Hozo3, Ambuj Kumar1,2, Benjamin Djulbegovic1,2,4*

1 Comparative Effectiveness Research, University of South Florida, Tampa, FL, United States of America, 2 Department of Internal Medicine, University of South Florida, Tampa, FL, United States of America, 3 Department of Mathematics, Indiana University of Northwest, Gary, IN, United States of America, 4 Departments of Hematology and Health Outcomes and Behavior, H. Lee Moffitt Cancer Center & Research Institute, Tampa, FL, United States of America

* [email protected]

Abstract Dual Processing Theories (DPT) assume that human cognition is governed by two distinct

types of processes typically referred to as type 1 (intuitive) and type 2 (deliberative). Based

on DPT we have derived a Dual Processing Model (DPM) to describe and explain therapeu-

tic medical decision-making. The DPMmodel indicates that doctors decide to treat when

treatment benefits outweigh its harms, which occurs when the probability of the disease is

greater than the so called “threshold probability” at which treatment benefits are equal to

treatment harms. Here we extend our work to include a wider class of decision problems

that involve diagnostic testing. We illustrate applicability of the proposed model in a typical

clinical scenario considering the management of a patient with prostate cancer. To that end,

we calculate and compare two types of decision-thresholds: one that adheres to expected

utility theory (EUT) and the second according to DPM. Our results showed that the deci-

sions to administer a diagnostic test could be better explained using the DPM threshold.

This is because such decisions depend on objective evidence of test/treatment benefits

and harms as well as type 1 cognition of benefits and harms, which are not considered

under EUT. Given that type 1 processes are unique to each decision-maker, this means

that the DPM threshold will vary among different individuals. We also showed that when

type 1 processes exclusively dominate decisions, ordering a diagnostic test does not affect

a decision; the decision is based on the assessment of benefits and harms of treatment.

These findings could explain variations in the treatment and diagnostic patterns docu-

mented in today’s clinical practice.

Introduction A paradigmatic decision-making dilemma faced by clinicians is whether to observe the patient without ordering a diagnostic test, order a diagnostic test and act according to the results of the test, or administer treatment without ordering a test. Typically, this decision relies on the prob- ability of disease and the relationship between the treatment’s harms and benefits. As described

PLOSONE | DOI:10.1371/journal.pone.0134800 August 5, 2015 1 / 16

OPEN ACCESS

Citation: Tsalatsanis A, Hozo I, Kumar A, Djulbegovic B (2015) Dual Processing Model for Medical Decision-Making: An Extension to Diagnostic Testing. PLoS ONE 10(8): e0134800. doi:10.1371/ journal.pone.0134800

Editor: Guy Brock, University of Louisville, UNITED STATES

Received: December 1, 2014

Accepted: July 14, 2015

Published: August 5, 2015

Copyright: © 2015 Tsalatsanis et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

Data Availability Statement: Data are presented in the manuscript.

Funding: This work is supported by the Department of Army grant #W81 XWH 09-2-0175. (PI: BD).

Competing Interests: The authors have declared that no competing interests exist.

later in this paper, the assessment of the likelihood of disease and the evaluation of treatment’s benefits and harms is often done intuitively, but this decision-making process can be formal- ized under the “threshold model”.

According to the threshold model [1,2], when faced with a choice of observing the patient, ordering a diagnostic test, or administering treatment, there is a probability of disease, also known as threshold probability, at which a decision maker is indifferent between any two choices (e.g. treating vs. ordering a test, or ordering a test vs. withholding treatment) [3–6]. Furthermore, decisions involving diagnostic testing rely on two probabilities of disease known as testing and treatment thresholds. Testing threshold relates to the decision about ordering a test vs. observing a patient and treatment threshold relates to the decision about administering treatment vs. ordering the diagnostic test. According to the threshold model [1,2], if the proba- bility of disease is smaller than the testing threshold, the test should be withheld. If the proba- bility of disease is above the treatment threshold, then treatment should be ordered without ordering a diagnostic test. The test should only be ordered if the estimated probability of the disease is between the testing and treatment thresholds (Fig 1).

The threshold model relies on expected utility theory (EUT) and it was formulated almost 4 decades ago[1,2]. EUT suggests that when choosing between different strategies, the decision maker should always select the strategy that leads to the outcome with the highest expected utility. It has been well documented, however, that EUT is routinely violated by decision-mak- ers [7–9]. These violations are typically attributed to the decision maker’s emotional, experien- tial or intuitive responses to decision choices that are different from the EUT derived expected utilities. Consequently, the main drawback of threshold model is its reliance on EUT as demon- strated in our recent empirical study [10].

The importance of non-EUT based cognitive processes has recently been highlighted by dual processing theories (DPT) of human reasoning and decision-making [11], which are increasingly accepted as the dominant explanation of how people make decisions [5,7,12–19]. DPT posits that human cognition is governed by two types of processes [11,19]: type 1 pro- cesses, which are intuitive, automatic, fast, narrative, experiential and affect-based, and type 2 processes, which are analytical, slow, verbal, deliberative and allow for abstract and hypotheti- cal thinking. Therefore, the EUT model cannot be seen as an adequate model of medical deci- sion-making.

Fig 1. Relation between the probability of disease and the threshold probabilities for testing and treatment (adopted from [3]).

doi:10.1371/journal.pone.0134800.g001

DPM for Medical Decision-Making

PLOS ONE | DOI:10.1371/journal.pone.0134800 August 5, 2015 2 / 16

To overcome the drawbacks of the EUT-based threshold model we recently developed a Dual Processing Model (DPM) [12], which is based on DPT. The DPM [12] incorporates regret to model type 1 processes and EUT to model type 2 processes. This is because regret is one of the key emotions that play a major role in medical-decision making [20–22]. Two main assumptions of DPM are that the extent of activation of type 1 processes is regulated by a parameter γ, and that when faced with a decision problem our initial responses tend to rely mostly on type 1 processes [23].

In our previous work [12] we demonstrated the applicability of the DPM-based threshold in a situation when no diagnostic test is available but a clinician has to make a decision whether to administer treatment or not. Here, we extend our work to include a wider class of medical decision-making problems that involve diagnostic testing.

Methods

Threshold models EUT threshold model. Most decision theories agree that decision-making depends on

evaluation of harms (losses) and benefits (gains) associated with a given decision strategy. The threshold model takes this into consideration by relating the threshold probability to benefit/ harms ratio. For example, the EUT threshold is calculated as:

pt;EUT ¼ 1

1þ BII HII

; forHII > 0

Where pt, EUT 2 [0,1] is the threshold probability i.e. the probability of disease at which we are indifferent between treatment vs. no treatment. BII � 0 is the net benefits of treatment defined as the difference in outcomes of treating and not treating a patient with disease, as realized by type 2 (denoted also as II in the equations) processes [12,24–29]. HII > 0 is the net harms due to treatment, defined as the difference in outcomes of not treating and treating the patients without disease as realized by type 2 processes[12,24–29]. Typically, the values of harms and benefits are obtained from the best available evidence found in the literature [12,24–29]. Note that for the validity of the pt,EUT equation, HII must take values greater than zero. This require- ment is clinically justifiable because in reality every treatment is associated with some harms.

Regret-based model. When treating a patient, a decision maker may face two types of regret: regret associated with failure to provide necessary treatment (regret of omission) and regret associated with administering harmful treatment (regret of commission) [12,21,22,30,31]. These two regrets are used to compute the regret based threshold probability as:

pt;RG ¼ 1

1þ BI HI

; forHI > 0

where pt,RG 2 [0,1] is the threshold probability at which a decision maker is indifferent between treating or not a patient. BI � 0 is the net benefits of treatment as realized by type 1 processes and computed here as regret of omission.HI > 0 is the net harms of treatment as realized by type 1 (denoted also as I in the equations) processes and computed here as regret of commis- sion. [12,21,22]. Both BI and HI values may be elicited using the Dual Analogue Scale described elsewhere [21,22]. As with HII,HI must take values greater than zero so that pt,RG is defined.

Dual Processing Model. DPM [12] assumes that the valuation of a risky choice is formed as the combination of type 1 and type 2 processes. To demonstrate, consider a clinical scenario (Fig 2) in which a decision maker is faced with a choice of treating (Rx) or not (NoRx) of a

DPM for Medical Decision-Making

PLOS ONE | DOI:10.1371/journal.pone.0134800 August 5, 2015 3 / 16

patient who has a disease with probability p. Each decision results in a specific outcome xi. For example, outcome x1 corresponds to the decision of treating a patient who had a disease and outcome x2 corresponds to the decision of treating a patient who did not have a disease. The parameters xmI

i � 0 and xmII i � 0 correspond to valuations of the outcome xi when the decision

maker employs type 1 and 2 processes respectively. Each outcome is also associated with type 1, UI,i � 0, and type 2, UII,i � 0, utilities.

Solving the decision tree in Fig 2, we derive the DPM threshold probability, pt 2 [0,1], or the probability at which we are indifferent between providing and withholding treatment, as [12]:

pt ¼ min ðpt;EUTÞ 1þ g 2ð1� gÞ

HI

HII

� � 1� BI

HI

� �� � ; 1

� � ; for g 2 ½0; 1� ð1Þ

The interaction between type 1 and type 2 processes is represented by the parameter γ 2 [0,1]. γ exemplifies the extent of activation of type 1 processes in the decision in such a way that when it is zero the decision-making processes is based on type 2 processes according to the EUT paradigm. As the value of γ increases, so does the involvement of type 1 processes in the decision. However, Eq 1 is not valid for values of γ = 1. In that case, based on the decision tree depicted in Fig 2 for γ = 1, the decision to treat or not depends solely on the explicit evaluation of harms and benefits based on type 1 processes (see also the Special Case in S1 Appendix). As a consequence, treatment should be administered only if benefits of treatment as assessed by type 1 processes outweigh harms of treatment.

Fig 2. Decision tree describing a typical scenario in which a physician is considering administering (Rx) / withholding treatment (NoRx) to/from his patient. xi represents an outcome; γ is the involvement of type 1 in the decision process; p is the probability of disease;UI,i is the utility of the outcome xi under type 1 process;UII,i is the utility of outcome xi under type 2 processes; The valuation of an outcome xi under type 1 is estimated as the regret associated with the outcome xi; the valuation of an outcome xi under type 2 is estimated as the utility of the outcome xi ([12] for details).

doi:10.1371/journal.pone.0134800.g002

DPM for Medical Decision-Making

PLOS ONE | DOI:10.1371/journal.pone.0134800 August 5, 2015 4 / 16

γ can be best visualized as the relative distance between the analytically derived threshold pt, EUT and the regret derived threshold pt,RG, or [32]:

g ¼ min jpt;EUT � pt;RGj

pt;EUT ; 1

( )

However, γ can be affected by many different mechanisms that characterize type 1 pro- cesses. Even though our model assumes a dominant role of regret, it does also incorporate other mechanisms of type 1 cognitive processes.

If a patient’s probability of disease is greater than the threshold probability then the decision maker favors treatment and withholds treatment otherwise. Eq 1 shows the impact of the extent of treatment harms and benefits on decisions and how they relate to the DPM and EUT thresholds. When the type 1 benefits of treatment as perceived by the decision maker are higher than its harms, the DPM threshold is always lower than the EUT threshold. Conversely, the DPM threshold is always greater than the EUT threshold if the type 1 harms of treatment are perceived to be higher than benefits. These changes in the threshold often lead to different choices than those predicted by the EUT and therefore may explain the violations of EUT in decision-making described extensively in literature [7,8,21,22,33,34].

DPMwith a diagnostic test. In many cases the use of a diagnostic test may assist the treat- ing physician in decreasing diagnostic uncertainty. However, obtaining diagnostic information may expose the patient to unnecessary risks [4] and therefore, a test should be ordered only when benefits of testing outweigh its risks [3].

Typically, deciding when to perform a diagnostic test relates to the assessment of the prior probability that a patient has a suspected disease [3]. If the probability of disease is very low or very high, then performing a diagnostic test may be unnecessary. As explained above, accord- ing to the threshold framework, there exists: 1. a probability at which we are indifferent between performing a diagnostic test and withholding treatment; and 2. a probability at which we are indifferent between performing a diagnostic test and administering treatment. These probabilities are formally known as the threshold probabilities for testing and they are decom- posed into 1. testing threshold (ptt) and 2. treatment threshold respectively (prx) [3]. Here we derive and present both threshold probabilities in terms of DPM [12].

We consider a generic scenario in clinical decision-making in which a decision maker is considering one of three strategies for the management of a patient’s condition (Fig 3). These strategies are: 1. do nothing (NoRx), 2. perform a diagnostic test (T), and 3. administer treat- ment (Rx). The patient may have a disease (D) with probability p. Each strategy results in an outcome xi, which is associated with a certain valuation, xmI

i � 0 when type 1 processes are involved and xmII

i � 0 when type 2 processes are employed. Each outcome has a utility UI,i� 0 for type 1 processes and UII,i � 0 for type 2 processes. As described earlier, valuation of out- comes under type 1 processes is performed using regret elicited using the Dual Visual Analogue Scale (DVAS) while valuation of outcomes under type 2 processes is based on EUT and the lat- est available evidence [12].

Solving the decision tree in Fig 3, we derive the following expected valuations for each strat- egy (detailed derivation is presented in the S1 Appendix):

VðRxÞ ¼ g 2 ðUI;2 � UI;4Þ þ ð1� gÞ½pUII;1 þ ð1� pÞUII;2�

VðNoRxÞ ¼ g 2 ðUI;3 � UI;1Þ þ ð1� gÞ½pUII;3 þ ð1� pÞUII;4�

DPM for Medical Decision-Making

PLOS ONE | DOI:10.1371/journal.pone.0134800 August 5, 2015 5 / 16

and

VðTÞ ¼ g 4 ðUI;2 � UI;4 þ UI;3 � UI;1Þ þ ð1� gÞ½pSUII;1 þ ð1� pÞð1� SpÞUII;2 þ pð1� SÞUII;3

þ ð1� pÞSpUII;4� � ðgHI;T þ ð1� gÞHII;TÞ

The notation for the expected valuations is as follows: UI,i� 0 and UII,i� 0 corresponds to the utilities of the xi outcome under type 1 and 2 processes respectively; p is the probability of disease; γ 2 [0,1] is the weight given to type 1 processes; S 2 [0,1] is the sensitivity of the diag- nostic test; Sp 2 [0,1] is the specificity of the diagnostic test; HI,T� 0 and HII,T� 0 denote the harms associated with the diagnostic test as perceived by type 1 and 2 processes respectively.

Threshold probabilities for testing Testing threshold. The testing threshold is the probability at which we are indifferent

between withholding treatment and ordering a diagnostic test [3]. Thus, working with the expected valuations for NoRx and T, V(NoRx) = V(T), and solving for the threshold probability we derive:

ptt ¼ min ptt;EUT 1þ g

4ð1� gÞð1� SpÞ 1þ 1 1�Sp

HII;T

HII

� HI

HII

1� BI

HI

� � þ 4

HI;T

HII

� �2 4

3 5; 1

8< :

9= ;; for g 2 ½0; 1� ð2Þ

Eq 2 is invalid for γ = 1. In that case, decision makers should always choose not to treat instead of testing i.e. ordering a diagnostic test does not contribute to the decision (see S1

Fig 3. Decision tree describing a typical scenario in which a physician is considering one the following three strategies: administering treatment (Rx); withholding treatment (NoRx); and performing a diagnostic test before deciding on treatment (Test). xi represents an outcome; γ is the involvement of type 1 in the decision process; p is the probability of disease;UI,i is the utility of the outcome xi under type 1 andUII,i is the utility of outcome xi under type 2 cognitive processes; HI,T denotes the harms of test as realized by type 1 andHII,T denotes the harms of test as realized by type 2 processes; P1 = pS+(1-p) (1-Sp); P11 = pS/P1; P12 = (1-p)(1-Sp)/P1; P2 = (1-p)Sp+p(1-s); P21 = p(1-s)/P2; P22 = (1-p)Sp/P2; S is the test’s sensitivity and Sp the test’s specificity. The valuation of an outcome xi under type 1 is estimated as the regret associated with the outcome xi; the valuation of an outcome xi under type 2 is estimated as the utility of the outcome xi.

doi:10.1371/journal.pone.0134800.g003

DPM for Medical Decision-Making

PLOS ONE | DOI:10.1371/journal.pone.0134800 August 5, 2015 6 / 16

Appendix Special Case section for details). This result is a function of type 1 processes, which do not have a role in calibration of probabilities, but treat each choice as “yes/no” outcome (see Fig 3). Indeed, the key role of a diagnostic test is to decrease uncertainty by increasing/decreas- ing probability of disease. When a decision-maker does not take this probability into account, then there is no sense in considering a diagnostic test.

If the probability of disease is greater than or equal to ptt the decision maker favors the diag- nostic test; otherwise, he/she prefers withholding treatment. The DPM testing threshold (Eq 2) is always higher than the analytically derived EUT testing threshold if the relationship between type 1 benefits and harms of treatment and harms of test is HI+4HI,T> BI. This relationship shows that if the harms of test and treatment are perceived greater than the benefits of treat- ment, the decision maker requires more certainty before testing. Conversely, the DPM testing threshold is always lower than the EUT testing threshold ifHI+4HI,T> BI, which demonstrates that the decision maker requires less certainty before testing. Note that the accuracy of the diag- nostic test (expressed in terms of sensitivity and specificity) does not affect this finding. Both ptt,EUT and ptt are undefined for the special case of Sp = 100%. However, as Sp!100%, ptt, EUT!1 and ptt!1. This finding demonstrates a decision maker’s aversion in providing diag- nostic testing, which is perceived to lead to more harms than benefits. Note that the values of γ,

Sp, S and HI HII

� affect the degree (“depth”) by which the DPM testing threshold ptt is greater/

lower than the classic EUT threshold ptt,EUT; however, they do not change the quality of the relationship.

Treatment threshold. The treatment threshold is the probability at which we are indiffer- ent between testing and administering treatment [3]. Working with the expected valuations of Rx and T, V(Rx) = V(T), and solving for the threshold probability we derive:

prx ¼ min prx;EUT 1þ g

4ð1� gÞSp 1� 1 Sp

HII;T

HII

� HI

HII

1� BI

HI

� � � 4

HI;T

HII

� �2 4

3 5; 1

8< :

9= ;; for g 2 ½0; 1� ð3Þ

Eq 3 is invalid for γ = 1. In that case, decision makers should always choose treating instead of testing (see S1 Appendix Special Case section for details). As outlined above, this result is a consequence of how type 1 processes work: by treating each choice as “yes/no” outcome there is no sense in taking diagnostic test probabilities into account (see Fig 3).

If the probability of disease is greater than or equal to prx the decision maker will choose to administer treatment; otherwise, he/she will prefer to perform a diagnostic test. The DPM treatment threshold (Eq 3) is always higher than the analytically derived EUT treatment threshold if the relationship between type 1 benefits and harms of treatment and harms of test is as follows:HI > BI+4HI,T. This relationship shows that if the decision maker assumes that the harms of treatment are higher than its benefits added to the harms of testing, then he requires more certainty before proceeding with treatment. Conversely, the DPM treatment threshold is always lower than the EUT treatment threshold ifHI < BI+4HI,T, which demon- strates that the decision maker requires less certainty before proceeding with treatment. As above, the test sensitivity and specificity does not affect this relationship. Both rules assume that the diagnostic test is objectively assessed (via type 2 functioning) to be less harmful than

the treatment, HII,T < HII, which is almost always the case. The values of γ, Sp, S, and HI HII

� affect the extent (“depth”) by which the dual threshold prx is greater/lower than the classic EUT threshold prx,EUT but does not change the quality of the relationship.

To summarize, for γ 2 [0,1) a decision maker will choose to perform a diagnostic test if the patient’s probability of disease is ptt � p< prx. The probabilities ptt and prx are functions of a

DPM for Medical Decision-Making

PLOS ONE | DOI:10.1371/journal.pone.0134800 August 5, 2015 7 / 16

decision maker’s attitudes towards treatment benefits and harms as well as harms of testing and they are derived using both type 1 and type 2 cognitive mechanisms. The probability of disease, p can be estimated by statistical evidence, and by the physician’s intuition and experi- ence. When γ = 1, the management choices are limited to treatment vs no treatment as testing results in the overall lower valuation in comparison with the other two alternatives. Therefore, the optimal decision is a function of the decision maker’s attitudes towards treatment benefits and harms as assessed by type 1 processes (see S1 Appendix Special Case section for details).

Case Study We will now demonstrate the applicability of the proposed method as it relates to decisions regarding performing prostate biopsy in a patient suspected of having prostate cancer. Con- cerns about prostate cancer may be raised by elevated values of the Prostate-specific Antigen (PSA) biomarker and/or by abnormalities found through Digital Rectal Examination (DRE). Further verification is obtained through a biopsy, which is currently the gold standard for diag- nosis of prostate cancer. During the prostate biopsy, several needles are inserted through the rectum wall into the areas of the prostate, where the abnormality is detected, to remove small amounts of tissue, which are later analyzed in the lab. The patient may experience discomfort, pain, bleeding, hematuria, infections, sepsis and vasovagal episodes as a result of the biopsy [35–40]. Table 1 summarizes the risks and benefits associated with prostate biopsy as reported in literature.

Contingent on the results of the biopsy, the treating urologist may choose to perform a radi- cal prostatectomy and surgically remove a part or all of the prostate gland. The goal of the pro- cedure is to cure or control the cancer. The procedure is performed either through an open surgery, where the surgeon makes a cut in the abdomen or between the testicles and the back passage, or laparoscopy, where the surgeon makes several small incisions in the pelvis. In both cases, the patient may experience major or minor complications after or during the surgery including heart problems, blood clots, blood loss, allergic reactions to anesthesia, infections, erectile dysfunction, urinary incontinence, damage to the urethra or the rectum [41–43]. How- ever, radical prostatectomy has statistically beneficial effect on patient’s survival compared to observation[44–46]. Table 2 summarizes the risks and benefits of radical prostatectomy as reported in literature.

For example consider the management strategies for a 66-year-old patient with elevated PSA and abnormal DRE: 1. do nothing (e.g. observe or wait for 6 months to repeat PSA and DRE), 2. perform a prostate biopsy and act accordingly, and 3. proceed directly with radical prostatectomy. Based on the evidence provided on Tables 1 and 2 we define the type 2 benefits and harms regarding radical prostatectomy as BII = 2.5% to 10% andHII = 0.4%. The harms associated with prostate biopsy are HII,T = 0.09%.

Table 1. Harms associated with cancer biopsy as reported in literature*.

Cancer biopsy

Harm (H) Size

Death (H) 0.09% [52]

Infections (H) 2–3% [37–39]

Hematuria (H) 50%-60% [35,36,38]

Discomfort, pain, bleeding, sepsis, vasovagal episodes (H) <10% [35–40]

* The data are related to transrectal ultrasound guided prostate biopsy.

doi:10.1371/journal.pone.0134800.t001

DPM for Medical Decision-Making

PLOS ONE | DOI:10.1371/journal.pone.0134800 August 5, 2015 8 / 16

For demonstration and simplification purposes, we will first describe how the decision dilemma described above would be solved relying solely on type 2 processes focusing on the most important harms and benefits i.e. those related to survival. An elaborate modification of the EUTmodel to include all harms and benefits reported in literature can also be implemented as in [24]. We assume that the values of sensitivity and specificity of the biopsy are equal to 86% and 94% respectively [47] (the highest reported values for biopsy guided by transrectal ultrasound (TRUS)). The EUT-based threshold probabilities, are derived by Eqs 2 and 3 (assuming γ = 0) and they are equal to ptt_EUT = 0% and prx_EUT = 21% (considering maximum benefit of treatment BII = 10%). The results show that a decision maker will accept biopsy and surgery at very low probabilities of prostate cancer: 0% for biopsy and 21% for surgery. That is, according to the EUT model we should perform biopsy at the slightest suspicion of prostate cancer (i.e., as long as it is greater than 0%!), and can recommend surgery at the estimated probability of prostate cancer> 21%! No physician (or, a patient) would agree with such rec- ommendations. The finding based on the EUT model also contradicts the influential, National Cancer Network (NCCN) expert guidelines for prostate cancer [48] which indicates that a prostate biopsy should be performed if the probability of prostate cancer exceeds 48% (indi- rectly computed by a Gleason score of 8 or higher for symptomatic patients which translates into 48% according to [49]).

Our finding of low threshold for the action may be attributed to the oversimplified assump- tion focusing only on mortality, which is rather small: harms (death) attributed to biopsy (0.09%) and prostatectomy (0.4%). If instead of death due to prostatectomy, we focus on erec- tile dysfunction (37%), which reflects the main concern of patients with 20 or more years of life expectancy, the threshold values increase considerably: ptt_EUT = 21% and prx_EUT = 49% (con- sidering maximum benefit of treatment BII = 10%). In this case a decision maker opts out of biopsy for probabilities less than 21% and requires more certainty for prostatectomy (49%). The problem, however, is how to integrate multiple outcomes in the EUT model, particularly since it is believed that the values people attach to different outcomes depend on the type 1 mechanisms, which processes the information on all benefits and harms in holistic fashion [4].

It is, therefore, necessary to arrive at decisions using cognitive mechanisms that employ both type 2 and type 1 processes. In our model, this is easily accomplished by increasing the value of γ, which reflects the extent of type 1 processes in the decision process. As a result both testing thresholds change. To compute the threshold values from Eqs 2 and 3 we need to elicit the decision maker’s preferences towards biopsy (HI,T) and towards prostatectomy (BI,HI). In contrast to the valuation of outcomes through EUT that entail inquiries for every harm and

Table 2. Benefits and harms of radical prostatectomy as reported in literature*.

Radical prostatectomy

Benefit (B) or Harm (H) Size

Survival (B) Absolute risk reduction: 2.5%- 10% [44–46]

Death (H) 0.4% [46]

Erectile dysfunction (H) 37% [46]

Urinary incontinence (H) 10.8% [46]

Heart problems (H), blood clots (H), blood loss (H), allergic reactions to anesthesia (H), infections (H), damage to the urethra or the rectum (H)

<10% [41–43]

* The data are related to radical prostatectomy performed as an open surgery or laparoscopic.

doi:10.1371/journal.pone.0134800.t002

DPM for Medical Decision-Making

PLOS ONE | DOI:10.1371/journal.pone.0134800 August 5, 2015 9 / 16

benefit individually, the type 1 processes valuate outcomes in a holistic manner by eliciting regrets of omission and commission using the DVAs [21,22]. Because type 1 processes are unique to each decision-maker, we expect that the DPM-based thresholds will vary among dif- ferent individuals.

Figs 4 and 5 graph the values of the EUT and DPM thresholds for testing as functions of the type 1 treatment benefit/harm ratio (BI/HI) for different values of type 1 harms of biopsy (HI,

T). Both figures are generated for maximum benefit of treatment (BII = 10%) however, Fig 4 assumes harms of treatment relate to survival (HII = 0.4%) and Fig 5 assumes harms of treat- ment relate to erectile dysfunction (HII = 37%).

Fig 4 demonstrates that as the harms of biopsy (HI,T) increase (Fig 4b, 4c and 4d), the decision maker will always choose a minimal risk and high benefit prostatectomy over the biopsy. This result is true for any type 1 treatment benefit/harm ratio (BI/HI). In addition, it is shown that the treatment threshold decreases dramatically as the type 1 benefits of prostatectomy outweigh its harms (BI> HI) and that there exists a benefits/harms ratio at which a decision maker would opt for prostatectomy at practically 0% probability of prostate cancer. These results appear to be dif- ferent than those based on the EUTmodel, which recommends biopsy to practically all patients and prostatectomy to the patients with probability of cancer greater than 21% regardless of the decision maker’s preferences towards biopsy and prostatectomy. Note the important result in Fig

Fig 4. EUT and DPM testing thresholds as functions of type 1 benefits/harms of prostatectomy ratio. The chart progression (Fig 4a–4d) shows the effect of increasing type 1 harms of biopsy on the values of testing thresholds. Unlike the EUT threshold, as harms of biopsy (HI,T) increase (Fig 4b, 4c and 4d), the DPM testing threshold increases to the maximum indicating that a decision maker will never choose a biopsy. When benefits of prostatectomy are higher than its harms (BI>HI), the decision maker opts for prostatectomy at practically 0% of disease. Note that the DPMmodel allows for the treatment threshold to be lower than the testing threshold. This is rationally not possible within the EUT framework, but has been observed in clinical practice. As an illustration consider the case where BI<HI. The DPM testing threshold (ptt) is always higher than the EUT testing threshold (ptt,EUT). This is because the DPM testing threshold considers the decision maker’s attitudes towards treatment according to which the benefits of treatment are higher than its harms (e.g. BII>HII). The same holds for the case of BII<HII, but only whenHI,T>0 (i.e. the diagnostic test is harmful) (Fig 4b, 4c and 4d). If HI,T = 0 and BII>HII (Fig 4a), the decision maker may choose test or treatment at the same probability of disease. Also, for most BI/HI, the DPM treatment threshold (prx) is lower than the EUT treatment threshold (prx,EUT). Again, this is because the decision maker values treatment benefits higher that its harms.

doi:10.1371/journal.pone.0134800.g004

DPM for Medical Decision-Making

PLOS ONE | DOI:10.1371/journal.pone.0134800 August 5, 2015 10 / 16

4: The treatment threshold is lower than the testing threshold, which is rationally not possible within the EUT framework, but has been observed in clinical practice [20].

The results shown in Fig 5 are far more complicated and fascinating since in this case we consider that the prostatectomy may cause erectile dysfunction in 37 out of 100 people (BII < HII). It would be logical to assume that a decision maker would always prefer to undergo a diag- nostic test with minimal harms before a harmful intervention. This behavior is demonstrated in Fig 5a, where it is assumed thatHI,T = 0. However, the prostate cancer probability levels at which a decision maker would opt in for either strategy changes based on the way the decision maker experiences type 1 benefits and harms of prostatectomy. For example, in Fig 5a, the decision maker will accept biopsy at a probability of prostate cancer greater than 40% (closer to NCCN guidelines than the EUT threshold) and prostatectomy for probability greater than 90% if he perceives type 1 harms of prostatectomy as greater than its benefits (e.g. HI = 8BI). If the decision maker perceives type 1 benefits of prostatectomy higher than its harms (e.g. BI = 8HI), he may tolerate biopsy for any probability of prostate cancer and accepts prostatectomy at a probability greater than 60%.

Furthermore, as the harms of biopsy (HI,T) increase (Fig 5b, 5c and 5d), there is a range of benefit/harms ratio above which the decision maker will prefer prostatectomy over biopsy. For example, when type 1 harms of biopsy are believed to be 10% (Fig 5b), a prostatectomy is pre- ferred to a biopsy (if type 1 benefits of prostatectomy are at least 2 times higher that its harms). Similarly, when HI,T = 20% (Fig 5c), a prostatectomy is preferred to a biopsy if its benefits are slightly higher than its harms. Once again, this is rationally not possible within the EUT frame- work[20], however, it is observed in current urological practice.

Fig 5. EUT and DPM testing thresholds as functions of type 1 benefits/harms of prostatectomy ratio. The chart progression (Fig 5a–5d) shows the effect of increasing type 1 harms of biopsy to the values of testing thresholds. The value of treatment threshold decreases as the ratio benefit/harms of prostatectomy increases (Fig 5a, 5b, 5c and 5d). The value of testing threshold also decreases as the ratio benefit/harms of prostatectomy increases but only when the harms of biopsy are zero (Fig 5a). If the decision maker perceives biopsy as harmful (Fig 5b, 5c and 5d) the testing threshold increases to the point that he will never choose biopsy. A prostatectomy becomes the preferred choice when BI > 2HI in Fig 5b; BI > HI in Fig 5c; BI > 0.8HI in Fig 5d.

doi:10.1371/journal.pone.0134800.g005

DPM for Medical Decision-Making

PLOS ONE | DOI:10.1371/journal.pone.0134800 August 5, 2015 11 / 16

Our case study clearly illustrates that the estimation of the exact values for each parameter, particularly those valuated under type 1 functioning, is not a simple exercise and reflect the complexity of real time decision making. It also demonstrates a variation in decisions made by different decision makers, a finding that may explain why people violate EUT.

Discussion In this article we described the derivation of a DPM for medical decision making to accommo- date decisions that involve diagnostic testing. Our model is based on Dual Processing Theories and assumes that human cognition relies on type 1 processes, which are intuitive and affect- based, as well as type 2 processes, which are analytical and deliberate processes.

Most medical decision making models rely on EUT and have failed to explain variations between the predicted versus the actual decisions people make. We hypothesize that this is because most existing models depend only on the analytical process of human cognition and ignore other experiential aspects of the decisions humans face[12,50].

Our DPMmodifies the EUT threshold model for decision making to incorporate influences by both modes of human cognition: intuitive, affect-based (type 1) and analytical processes (type 2). The derived expressions (Eqs 2 and 3) show that the DPM thresholds modify the EUT thresholds based on the way decision makers evaluate trade-offs of treatment. For example, the DPM-based testing threshold is always higher than the EUT-based threshold when harms of treatment and of test are assessed by type 1 processes to be higher than benefits of treatment. On the other hand, the DPM-based treatment threshold is always lower than the EUT-based threshold when harms of treatment are perceived by the type 1 processes to be lower than treat- ment benefits and biopsy harms. As described in the Methods section, the test sensitivity and specificity does not affect this relationship. The importance of our findings is best seen in the context of the current attempts to curb waste associated with over-testing. The American Board of Internal Medicine’s (ABIM) nine specialty societies representing 374,000 physicians developed a list of each specialty’s ‘Top Five’ inappropriately prescribed diagnostic tests in order to improve care by eliminating unnecessary tests and procedures [51]. For example, one typical recommendation reads, “Don’t order annual electrocardiograms (EKGs) or any other cardiac screening for low-risk patients without symptoms. False-positive tests are likely to lead to harm through unnecessary invasive procedures, over-treatment and misdiagnosis.” The problem with this guideline is how to determine how low is “low” (in terms of “low-risk”) and how likely is “likely” (in terms of false positives)? That is, at which threshold probability the test should actually be ordered? As illustrated in this paper, because the test characteristics do not affect the thresholds, we do not need to worry about false-positives (or, false-negatives for that matter). What matters is 1) objective data on treatment benefits and harms, and 2) how the decision-maker perceives these data (via type 1 processes). Therefore, a solution to the cur- rent health care waste could be to continue emphasizing the need for reliable, evidence-based resources and to highlight the importance of cognitive mechanisms in the way we process the information we access through the literature and collect during a clinical encounter. It has been argued that mindful awareness of type 1 and type 2 processes [15] may help us improve our decision-making processes. Our model provides the salient outline of these processes and how these processes can be effectively approached in clinical practice and education.

We demonstrated the applicability of our approach in a hypothetical case study in which a decision maker is considering radical prostatectomy for a patient who has elevated PSA and abnormal DRE. As the decision maker is uncertain whether a radical prostatectomy is the appropriate action for the particular patient, he also considers a prostate biopsy. Our results demonstrated the inability of EUT to model the preferences and attitudes of individual decision

DPM for Medical Decision-Making

PLOS ONE | DOI:10.1371/journal.pone.0134800 August 5, 2015 12 / 16

makers towards treatment and diagnostic testing. On the other hand, the DPM threshold pro- vides a convincing explanation as to why treatment decisions vary between decision makers. We posit that this variation is contingent on the extent of activation of type 1 processes. Our main point here is that type 2 processes (as adhered to the EUT model) will always produce the same results, while it is type 1 processes that are unique to each decision-maker. In turn, it is the extent of activation of type 1 processes that can explain excessive ordering of diagnostic tests as well as overall variations in the treatment and diagnostic patterns documented in today’s clinical practice.

Our model has limitations. Even though it is a simple mathematical expression, its applica- tion is challenging since many of the model parameters are not easily elicited. Our suggestion, which was implemented in this paper, is to use data from published literature to valuate out- comes under type 2 processes and the Dual Visual Analogue scales developed in [21,22] to val- uate outcomes under type 1 processes. In fact, we have shown that it is possible to elicit these values [32] although we have not yet done it in the context of the diagnostic setting. Our next step is to perform decision-making experiments initially in simulated environments, through hypothetical scenarios, and later in real clinical environments. In both cases, we will compare the decisions predicted by the DPMmodel to the actual decisions physicians make.

Throughout this article we avoided assigning a clear role to the decision maker. We believe that our methodology can be used by physicians and/or by patients. It is our position, however, that medical decision-making is shared between physicians and their patients. In such setting physicians recommend alternative treatment strategies with their associated harms and bene- fits and the patients eventually agree with the recommendation. We envision our methodology as a part of a computerized decision support system operated by the physician to elicit the patient’s preferences towards alternative forms of treatment.

Conclusions We have extended the recently derived DPM for medical decision making to include a diagnos- tic testing. Our model has the potential to explain the discrepancies found between optimal and actual actions. Because it captures the salient elements of medical decision-making via few parameters, our model has offered an important didactic value for medical education. Future research involves testing our model in a simulated environment with a wide variety of health- care professionals.

Supporting Information S1 Appendix. Detailed derivations of Dual Processing thresholds. (DOCX)

Acknowledgments This work is supported by the Department of Army grant #W81 XWH 09-2-0175. (PI: BD)

Author Contributions Conceived and designed the experiments: AT IH BD. Performed the experiments: AT IH BD. Analyzed the data: AT IH BD. Contributed reagents/materials/analysis tools: AT IH AK BD. Wrote the paper: AT IH BD.

DPM for Medical Decision-Making

PLOS ONE | DOI:10.1371/journal.pone.0134800 August 5, 2015 13 / 16

References 1. Pauker S, Kassirer J (1975) Therapeutic decision making: a cost benefit analysis. N Engl J Med 293:

229–234. PMID: 1143303

2. Pauker SG, Kassirer J (1980) The threshold approach to clinical decision making. N Engl J Med 302: 1109–1117. PMID: 7366635

3. Sox HC, Higgins MC (1988) Medical decision making: Amer College of Physicians.

4. Hunink MGM, Glasziou PP, Siegel JE, Weeks JC, Pliskin JS, Elstein AS, et al. (2001) Decision making in health and medicine: integrating evidence and values: Cambridge University Press.

5. Kahneman D (2003) Maps of bounded rationality: psychology for behavioral economics. American Eco- nomic Review 93: 1449–1475.

6. Kahnemen D (2011) Thinking fast and slow. New York: Farrar, Straus and Giroux.

7. Baron J (2000) Thinking and deciding: Cambridge University Press.

8. Bell DE, Raiffa H, Tversky A (1988) Decision making: Descriptive, normative, and prescriptive interac- tions: Cambridge University Press.

9. Hastie R, Dawes RM (2009) Rational choice in an uncertain world: The psychology of judgment and decision making: SAGE Publications, Incorporated.

10. Djulbegovic B, Elqayam S, Reljic T, Hozo I, Miladinovic B, Tsalatsanis A, et al. (2014) How do physi- cians decide to treat: an empirical evaluation of the threshold model. BMCMedical Informatics and Decision Making 14: 47. doi: 10.1186/1472-6947-14-47 PMID: 24903517

11. Evans JSBT (2003) In two minds: dual-process accounts of reasoning. Trends in cognitive sciences 7: 454–459. PMID: 14550493

12. Djulbegovic B, Hozo I, Beckstead J, Tsalatsanis A, Pauker S (2012) Dual processing model of medical decision-making. BMCMedical Informatics and Decision Making 12: 94. doi: 10.1186/1472-6947-12- 94 PMID: 22943520

13. Kahneman D, Slovic P, Tversky A (2005) Judgement under uncertainty: heuristics and biases. New York: Cambridge University Press.

14. Croskerry P (2009) A universal model of diagnostic reasoning. Acad Med 84: 1022–1028. doi: 10. 1097/ACM.0b013e3181ace703 PMID: 19638766

15. Croskerry P (2013) FromMindless to Mindful Practice—Cognitive Bias and Clinical Decision Making. The New England journal of medicine 368: 2445–2448. doi: 10.1056/NEJMp1303712 PMID: 23802513

16. Croskerry P, Nimmo GR (2011) Better clinical decision making and reducing diagnostic error. J R Coll Physicians Edinb 41: 155–162. doi: 10.4997/JRCPE.2011.208 PMID: 21677922

17. Evans JSTBT (2007) Hypothethical thinking. Dual processes in reasoning and judgement. New York: Psychology Press: Taylor and Francis Group.

18. Slovic P, Finucane ML, Peters E, MacGregor DG (2004) Risk as analysis and risk as feelings: some thoughts about affect, reason, risk, and rationality. Risk Anal 24: 311–322. PMID: 15078302

19. Stanovich KE (1999) Who is rational?: Studies of individual differences in reasoning: Lawrence Erlbaum.

20. Hozo I, Djulbegovic B (2008) When is diagnostic testing inappropriate or irrational? Acceptable regret approach. Medical Decision Making 28: 540–553. doi: 10.1177/0272989X08315249 PMID: 18480041

21. Tsalatsanis A, Barnes LE, Hozo I, Djulbegovic B (2011) Extensions to Regret-based Decision Curve Analysis: An application to hospice referral for terminal patients. BMCMedical Informatics and Decision Making 11: 77. doi: 10.1186/1472-6947-11-77 PMID: 22196308

22. Tsalatsanis A, Hozo I, Vickers A, Djulbegovic B (2010) A regret theory approach to decision curve anal- ysis: A novel method for eliciting decision makers' preferences and decision-making. BMCMedical Informatics and Decision Making 10: 51. doi: 10.1186/1472-6947-10-51 PMID: 20846413

23. Thompson VA, ProwseTurner JA, Pennycook G (2011) Intuition, reason, and metacognition. Cognitive Psychology 63: 107–140. doi: 10.1016/j.cogpsych.2011.06.001 PMID: 21798215

24. Djulbegovic B, Hozo I, Lyman GH (2000) Linking evidence-based medicine therapeutic summary mea- sures to clinical decision analysis. MedGenMed 2: E6.

25. Cornelissen JJ, van PuttenWL, Verdonck LF, Theobald M, Jacky E, Daenen SM, et al. (2007) Results of a HOVON/SAKK donor versus no-donor analysis of myeloablative HLA-identical sibling stem cell transplantation in first remission acute myeloid leukemia in young and middle-aged adults: benefits for whom? Blood 109: 3658–3666. PMID: 17213292

DPM for Medical Decision-Making

PLOS ONE | DOI:10.1371/journal.pone.0134800 August 5, 2015 14 / 16

26. Kearon C, Akl EA, Comerota AJ, Prandoni P, Bounameaux H, Goldhaber SZ, et al. (2012) Antithrombo- tic therapy for VTE disease: Antithrombotic Therapy and Prevention of Thrombosis, 9th ed: American College of Chest Physicians Evidence-Based Clinical Practice Guidelines. Chest 141: e419S–494S. doi: 10.1378/chest.11-2301 PMID: 22315268

27. Koreth J, Schlenk R, Kopecky KJ, Honda S, Sierra J, Djulbegovic BJ, et al. (2009) Allogeneic stem cell transplantation for acute myeloid leukemia in first complete remission: systematic review and meta- analysis of prospective clinical trials. JAMA 301: 2349–2361. doi: 10.1001/jama.2009.813 PMID: 19509382

28. Linkins L-A, Choi PT, Douketis JD (2003) Clinical Impact of Bleeding in Patients Taking Oral Anticoagu- lant Therapy for Venous Thromboembolism: A Meta-Analysis. Ann Intern Med 139: 893–900. PMID: 14644891

29. Segal JB, Eng J, Jenckes MW, Tamariz LJ, Bolger DT, Krishnan JA, et al. (2003) Diagnosis and Treat- ment of Deep Venous Thrombosis and Pulmonary Embolism. Washington, DC: Agency for Healthcare Research and Quality, U.S. Department of Health and Human Services.

30. Djulbegovic B, Hozo I, Schwartz A, McMasters K (1999) Acceptable regret in medical decision making. Med Hypotheses 53: 253–259. PMID: 10580533

31. Djulbegovic B, Ende J, HammRM, Mayrhofer T, Hozo I, Pauker SG (2015) When is rational to order a diagnostic test, or prescribe treatment: the threshold model as an explanation of practice variation. European journal of clinical investigation 45: 485–493. doi: 10.1111/eci.12421 PMID: 25675907

32. Djulbegovic B, Elqayam S, Reljic T, Hozo I, Tsalatsanis A, Kumar A, et al. (2013) Empirical evaluation of the threshold model. BMCMed Inform Decis Mak under review.

33. Dawes RM, Kagan J (1988) Rational choice in an uncertain world: Harcourt Brace Jovanovich New York.

34. Hozo I, Djulbegovic B (2009) Will Insistence on Practicing Medicine According to Expected Utility The- ory Lead to an Increase in Diagnostic Testing? Reply to DeKay's Commentary: Physicians' Anticipated Regret and Diagnostic Testing. Medical Decision Making 29: 320–324.

35. De Jesus C, Corrêa LA, Padovani CR (2006) Complications and risk factors in transrectal ultrasound- guided prostate biopsies. Sao Paulo Med J 124: 198–202. PMID: 17086300

36. Djavan B, Waldert M, Zlotta A, Dobronski P, Seitz C, Remzi M, et al. (2001) Safety and morbidity of first and repeat transrectal ultrasound guided prostate needle biopsies: results of a prospective European prostate cancer detection study. The Journal of urology 166: 856–860. PMID: 11490233

37. Lee G, Attar K, Laniado M, Karim O (2006) Safety and detailed patterns of morbidity of transrectal ultra- sound guided needle biopsy of prostate in a urologist-led unit. International urology and nephrology 38: 281–285. PMID: 16868698

38. Rodriguez LV, Terris MK (1998) Risks and complications of transrectal ultrasound guided prostate nee- dle biopsy: a prospective study and review of the literature. The Journal of urology 160: 2115–2120. PMID: 9817335

39. Wagenlehner FME, Van Oostrum E, Tenke P, Tandogdu Z, Çek M, Grabe M, et al. (2012) Infective complications after prostate biopsy: outcome of the Global Prevalence Study of Infections in Urology (GPIU) 2010 and 2011, a prospective multinational multicentre prostate biopsy study. European urology.

40. Sieber PR, Rommel F, Theodoran CG, Hong RD, Del Terzo MA (2007) Contemporary prostate biopsy complication rates in community-based urology practice. Urology 70: 498–500. PMID: 17905105

41. Guillonneau B, Rozet F, Cathelineau X, Lay F, Barret E, Doublet JD, et al. (2002) Perioperative compli- cations of laparoscopic radical prostatectomy: the Montsouris 3-year experience. The Journal of urol- ogy 167: 51. PMID: 11743274

42. Fowler FJ, Barry MJ, Lu-Yao G, Roman A, Wasson J, Wennberg JE (1993) Patient-re ported complica- tions and follow-up treatment after radical prostatectomy: The national medicare experience: 1988– 1990 (updated June 1993). Urology 42: 622–628. PMID: 8256394

43. Hu JC, Nelson RA, Wilson TG, Kawachi MH, Ramin SA, Lau C, et al. (2006) Perioperative complica- tions of laparoscopic and robotic assisted laparoscopic radical prostatectomy. The Journal of urology 175: 541–546. PMID: 16406991

44. Abdollah F, Sun M, Schmitges J, Tian Z, Jeldres C, Briganti A, et al. (2011) Cancer-specific and other- cause mortality after radical prostatectomy versus observation in patients with prostate cancer: compet- ing-risks analysis of a large North American population-based cohort. European urology 60: 920–930. doi: 10.1016/j.eururo.2011.06.039 PMID: 21741762

45. Bill-Axelson A, Holmberg L, Ruutu M, Häggman M, Andersson SO, Bratell S, et al. (2005) Radical pros- tatectomy versus watchful waiting in early prostate cancer. New England Journal of Medicine 352: 1977–1984. PMID: 15888698

DPM for Medical Decision-Making

PLOS ONE | DOI:10.1371/journal.pone.0134800 August 5, 2015 15 / 16

46. Wilt TJ, Brawer MK, Jones KM, Barry MJ, AronsonWJ, Fox S, et al. (2012) Radical prostatectomy ver- sus observation for localized prostate cancer. New England Journal of Medicine 367: 203–213. doi: 10.1056/NEJMoa1113162 PMID: 22808955

47. Purohit RS, Shinohara K, Meng MV, Carroll PR (2003) Imaging clinically localized prostate cancer. The Urologic clinics of North America 30: 279. PMID: 12735504

48. National Comprehensive Cancer Network (2013) NCCNClinical Practice Guidelines in Oncology, Prostate Cancer, version 1.2013.

49. Andrén O, Fall K, Franzén L, Andersson SO, Johansson JE, Rubin MA (2006) How well does the Glea- son score predict prostate cancer death? A 20-year followup of a population based cohort in Sweden. The Journal of urology 175: 1337–1340. PMID: 16515993

50. Mukherjee K (2010) A dual systemmodel of preferences under risk. Psychological Review 177: 243– 255.

51. American Board of Internal Medicine.

52. Nam RK, Saskin R, Lee Y, Liu Y, Law C, Klotz LH, et al. (2010) Increasing hospital admission rates for urological complications after transrectal ultrasound guided prostate biopsy. The Journal of urology 183: 963–969. doi: 10.1016/j.juro.2009.11.043 PMID: 20089283

DPM for Medical Decision-Making

PLOS ONE | DOI:10.1371/journal.pone.0134800 August 5, 2015 16 / 16

Djulbegovic et al. BMC Medical Informatics and Decision Making 2012, 12:94 http://www.biomedcentral.com/1472-6947/12/94

RESEARCH ARTICLE Open Access

Dual processing model of medical decision-making Benjamin Djulbegovic1,2,3,7*, Iztok Hozo4, Jason Beckstead5, Athanasios Tsalatsanis1,2 and Stephen G Pauker6

Abstract

Background: Dual processing theory of human cognition postulates that reasoning and decision-making can be described as a function of both an intuitive, experiential, affective system (system I) and/or an analytical, deliberative (system II) processing system. To date no formal descriptive model of medical decision-making based on dual processing theory has been developed. Here we postulate such a model and apply it to a common clinical situation: whether treatment should be administered to the patient who may or may not have a disease.

Methods: We developed a mathematical model in which we linked a recently proposed descriptive psychological model of cognition with the threshold model of medical decision-making and show how this approach can be used to better understand decision-making at the bedside and explain the widespread variation in treatments observed in clinical practice.

Results: We show that physician’s beliefs about whether to treat at higher (lower) probability levels compared to the prescriptive therapeutic thresholds obtained via system II processing is moderated by system I and the ratio of benefit and harms as evaluated by both system I and II. Under some conditions, the system I decision maker’s threshold may dramatically drop below the expected utility threshold derived by system II. This can explain the overtreatment often seen in the contemporary practice. The opposite can also occur as in the situations where empirical evidence is considered unreliable, or when cognitive processes of decision-makers are biased through recent experience: the threshold will increase relative to the normative threshold value derived via system II using expected utility threshold. This inclination for the higher diagnostic certainty may, in turn, explain undertreatment that is also documented in the current medical practice.

Conclusions: We have developed the first dual processing model of medical decision-making that has potential to enrich the current medical decision-making field, which is still to the large extent dominated by expected utility theory. The model also provides a platform for reconciling two groups of competing dual processing theories (parallel competitive with default-interventionalist theories).

Background Dual processing theory is currently widely accepted as a dominant explanation of cognitive processes that charac- terizes human decision-making [1-9]. It assumes that cognitive processes are governed by so called system I (which is intuitive, automatic, fast, narrative, experiential and affect-based) and system II (which is analytical, slow, verbal, deliberative and logical) [1-10]. The vast majority

* Correspondence: [email protected] 1Center for Evidence-based Medicine and Health Outcomes Research, Tampa, FL, USA 2Department of Internal Medicine, Division of Evidence-based Medicine and Health Outcomes Research University of South Florida, Tampa, FL, USA Full list of author information is available at the end of the article

© 2012 Djulbegovic et al.; licensee BioMed Ce Creative Commons Attribution License (http:/ distribution, and reproduction in any medium

of existing models of decision-making including expected- utility theory, prospect theory, and their variants assume a single system of human thought [11]. Recently, formal models for integrating system I with system II models have been developed [3,11]. One such attractive model-Dual System Model (DSM)- has been devel- oped by Mukherjee [11]. Here, we extend Mukherjee’s DSM model to medical field (DSM-M) by linking it to the threshold concept of decision-making [12-15]. We also take into account decision regret, as an ex- emplar of affect or emotion that is involved in system I decision-making [2], and which is of particular rele- vance to medical decision-making [16-19]. Regret was also selected for use in our model because any

ntral Ltd. This is an Open Access article distributed under the terms of the /creativecommons.org/licenses/by/2.0), which permits unrestricted use, , provided the original work is properly cited.

Djulbegovic et al. BMC Medical Informatics and Decision Making 2012, 12:94 Page 2 of 13 http://www.biomedcentral.com/1472-6947/12/94

“theory of choice that completely ignores feeling such as the pain of losses and the regret of mistakes is not only descriptively unrealistic but also might lead to prescriptions that do not maximize the utility of out- comes as they are actually experienced” [1,20]. As more than 30% of medical interventions are cur-

rently not appropriately applied, mostly as over – or- undertreatment [21-23], we illustrate how the DSM-M model may be used to explain the practice patterns seen in the current medical practice. Our DSM-M model is primarily an attempt to describe how medical decisions are made. As a descriptive model its validation will re- quire comparing its outputs to actual choices made by patients and clinicians and their verbalized reactions to our model. We conclude the paper by providing some testable empirical predictions.

Methods A dual system model Building on the previous empirical research, which has convincingly showed that human cognition is deter- mined by both system I and system II processes [1,2,5,24,25]. Mukherjee recently developed a formal mathematical model, which assumes parallel functioning by both systems, while the final decision is a weighted combination of the valuations from both systems based on the value maximization paradigm (Figure 1) [11]. (NB. In this paper we employ terms system I and system II as popularized by Kahneman [1,2] although some authors prefer to talk about type 1 and 2 processing as it is almost certain that human cognition is not organized in distinctly separated physical systems [5,26,27]). Mukherjee’s dual system model (DSM) assumes that

evaluation of risky choice (C) is formed by the combined input of system I and system II into a single value and can be formulated as follows:

E Cð Þ ¼ γVI Cð Þ þ 1� γð ÞVII Cð Þ ¼ γ

1 n

X i

VI xið Þ þ 1� γð Þk X i

piVII xið Þ ð1Þ

Where C represents a decision-making situation (“choice”), n – number of outcomes, pi – probability of the ith outcome, xi, of the selected choice. VI represents

Decision/choice underuncertainty

(C)

Valuation by intuitiv affective cognitive s

Valuation by delibera analytical/logical cog (System II); VII (C)

(System I); VI (C)

Figure 1 Model of decision-making using dual processing cognitive p

valuation of decision under autonomous, intuitive, sys- tem I-based mode of decision-making and VII, which can be a utility function, represents valuation under a deliberative, rule-based, system II mode of decision- making. k-is a scaling constant, and γ [0 to 1] is the weight given to system I and can be interpreted as the relative extent of involvement of system I in the decision-making process [11]. System II is not split into two subsystems advocated by some [5], but is assumed to adhere to the rationality criteria of expected utility theory (EUT) as also advocated by modern decision sci- ence [11,28]. γ is assumed to be influenced by a number of processes that determine system I functioning. Mukherjee emphasized the following factors as the im- portant determinants of system I functioning [11]: indi- vidual decision-making and thinking predispositions [ranging from expected utility theory (EUT) “maximi- zers” to system I driven “satisficing” with no regard to probabilities but with editing or selection of outcomes of interest] [29], affective nature of outcomes (the higher the affective nature of outcomes, the higher is γ) and framing and construing the decision-making task (deci- sions for the self will likely have higher γ, as well as deci- sion problems that are contextualized and those requiring immediate resolution or are made under time pressure; the last four describe circumstances character- istic of medical decision-making). Easily available infor- mation, our previous experience, the way in which information is processed (verbatim vs. getting the “gist” of it) [30] as well as memory limitations [31] are also expected to affect γ. γ is, therefore, expected to be higher when information about probabilities and out- comes are ambiguous or not readily available, or when a very severe negative prior outcome is recalled [2,32,33]. On the other hand, when such data are available their joint evaluation by system II will reduce γ [11]. In gen- eral, the factors that define the process of system I can be classified under 4 major categories: a) affect, b) evolu- tionary hard-wired processes, responsible for automatic responses to potential danger in such a way that system I typically gives higher weight to potentially false posi- tives than to false negatives (i.e. humans are cognitively more ready to wrongly accept the signal of potential harms than one that carries the potential of benefit),

e, experiential, ystem

tive, reflective nitive system

Final choice VC=VI(C)+VII(C)

rocesses (after Mukherjee [11]).

Djulbegovic et al. BMC Medical Informatics and Decision Making 2012, 12:94 Page 3 of 13 http://www.biomedcentral.com/1472-6947/12/94

(c) over-learned processes from system II that have been relegated to system I (such as the effect of intensive training resulting in the use of heuristics, or “rules of thumb” or practice guidelines as one of the effort-saving cognitive strategies. NB although guidelines may be the products of analytic system II processes their applica- tions tends to be a system I process.), and (d) the effects of tacit learning [5]. Mukherjee’s DSM model draws upon empirical evidence

demonstrating that decision-makers in an affect-rich con- text are generally sensitive only to the presence or absence of stimuli, while in affect-poor contexts they rely on sys- tem II to assess the magnitude of stimuli (and probabil- ities) [24]. Hence, the salient feature of the model is that that system I recognizes outcomes only as being possible or, not. Every outcome that remains under consideration gets equal weight in system I. On the other hand, system II recognizes probabilities linearly without distortions, according to the expected utility paradigm. As a result, dual valuation processing often generates

instances where subjective valuations are greater at lower stimulus magnitudes (i.e. when decision-making relies on feeling, or evolutionary hard-wired processes such as when the signal may present danger) while ra- tional calculation produces greater value at high magni- tudes [11]. DSM is capable of explaining a number of the phenomena that characterize human decision- making such as a) violation of nontransparent stochastic dominance, b) fourfold pattern of risk attitude, c) ambi- guity aversion, d) common consequences effect, e) com- mon ratio effect, f ) isolation effect, g) and coalescing and event-splitting effect [11]. Under the realistic assumption that outcomes are posi-

tive (i.e., utilities >0, which is particularly applicable to medical setting) and power value functions, VI xð Þ ¼ xmI ,

System I

System II

System II

System I

Figure 2 Dual processing model of decision-making as applied to a c (D+) or not. The patient may or may not have a disease (probability p). Re competing treatment alternative may include Rx or NoRx). Rg- regret.

and VII xð Þ ¼ x for system I and system II, respectively, DSM can be re-written as:

V Cð Þ ¼ γ 1 n

X i

xmI i þ 1� γð Þk

X i

pixi ð2Þ

where 0 <mI ≤ 1 Note that xmI i satisfies risk aversion for

gains and risk seeking for losses and that the term for system II pixi is linear without risk distortions. As noted by Mukherjee [11], the estimation of the

parameters in Equation 2) is a measurement exercise, which needs to be evaluated in the future empirical re- search. Consequently, the functions VII(x) and VI(x) could be changed, depending on the decision-making setting and decision-maker’s goals. Similarly, parameter m may not be the same for all outcomes.

Modification of DSM for medical decision-making We will consider a typical situation in clinical decision- making where a doctor has to choose treatment (Rx) vs. no treatment (NoRx) for disease (D) which is present with the probability p. [Note than NoRx represents a competing treatment alternative and may include a dif- ferent treatment (Rx2)] [12,34]. Each decision results in outcomes that have a certain value, xi. The model is shown in the Figure 2. As noted above, the system I recognizes outcomes only as being possible (or not), and is thus insensitive to exact probabilities. Every outcome with non-zero probability gets equal weight in system I. Hence, in a two-alternative choice, each probability is equal to 0.5 under system I. System II recognizes prob- abilities without distortions, as would be expected according to EUT. We posit that among the emotions that can influence

valuation of outcomes in system I processing, regret plays

linical dilemma whether to treat (Rx) the patient with disease gret is assumed to operate at the level of system I only. (Note that

Djulbegovic et al. BMC Medical Informatics and Decision Making 2012, 12:94 Page 4 of 13 http://www.biomedcentral.com/1472-6947/12/94

an important role [1,2], while system II processes are domi- nated by rational, analytical deliberations according to EUT [11]. We can define regret (Rg) as the difference (loss) in the utilities of the outcome of the action taken and that of the action we should have taken, in retrospect [16-19,35] but operating at the system I level only (see Figure 2). Hence, we have the following value functions (see

Additional file 1: Appendix for detailed derivation):

VI Rx;Dþð Þ ¼ Rg Rx;Dþ½ � ¼ 0; VI NoRx;Dþð Þ ¼ Rg NoRx;Dþ½ � ¼ xmI

3 � xmI 1 ;

VII Rx; Dþð Þ ¼ x1 ; VII NoRx; Dþð Þ ¼ x3;

VI Rx;D�ð Þ ¼ Rg Rx;D�½ � ¼ xmI 2 � xmI

4 VI NoRx; D�ð Þ ¼ Rg NoRx; D�½ � ¼ 0;

VII Rx; D�ð Þ ¼ x2; VII NoRx; D�ð Þ ¼ x4;

Overall valuation of decision to treat (Rx) is equal to:

V Rxð Þ ¼ γ

2 VA Rx;Dþð Þ þ VA Rx; D�ð Þð Þ

þ 1� γð ÞkðpVDðRx;DþÞ þ 1� pð ÞVD Rx; D�ð ÞÞ

¼ γ

2 xmA 2 � xmA

4ð Þ þ 1� γð Þk px1 þ 1� pð Þx2½ � And

V NoRxð Þ ¼ γ

2 VA NoRx;Dþð Þ þ VA NoRx; D�ð Þð Þ

þ 1� γð ÞkðpVDðNoRx;DþÞ þ 1� pð ÞVD NoRx; D�ð ÞÞ

¼ γ

2 xmA 3 � xmA

1ð Þ þ 1� γð Þk px3 þ 1� pð Þx4½ �

The difference in the outcomes of treating and not treat- ing patient with disease are equal to the net benefit of treatment (B) [13,14,36]; the difference in outcomes of not treating and treating those patients without disease is defined as net harms (H) [13,14,36]. Note that benefits and harms can be expressed in the various units (such as survival, mortality, morbidity, costs, etc.) and can be for- mulated both as utilities and disutilities [13,14,36]. As explained above, we further assume that valuation of net benefits and net harms by system I differs from system II. Hence, under system II, we replace net benefit and net harms using EUT definitions:BII ¼ x1 � x3 and net harms HII ¼ x4 � x2 . Under system I, we define BI ¼ xmI

1 � xmI 3 ,

and HI ¼ xmI 4 � xmI

2 . Solving for p (the probability of

disease at which we are indifferent between Rx and NoRx), we obtain: (Equation 3)

pt ¼ p ¼ 1� γð ÞkHII � γ 2 BI �HI½ �

1� γð Þk BII þ HII½ � ¼ 1

1þ BII HII

� γ

2k 1� γð Þ BI � HI

BII þ HII

¼ 1

1þ BII HII

! 1þ γ

2 1� γð Þ HI

HII

� � 1� BI

HI

� �� �

¼ pt EUTð Þð Þ 1þ γ

2 1� γð Þ HI

HII

� � 1� BI

HI

� �� � ð3Þ

This means that if the probability of disease is above pt the decision-maker favors treatment; otherwise, a compet- ing management alternative (such as “No Treatment”) represents the optimal treatment strategy. Note that k can be typically set at 1, as we do it here. Also note that the first part of equation is equivalent to the threshold expres- sion described in EUT framework [13,14,36]; the second expression modifies system II’s EUT-based decision-mak- ing process in such a way that if benefits are experienced higher than harms, the threshold probability is always lower than EUT threshold. However, if a decision-maker experiences HI>BI, the threshold probability is always higher than the EUT threshold (see below for discussion in the context of medical example). Note that γ and the ratio HI

HII only contribute to the extent of magnitude the

dual threshold is above or below the classic EUT thresh- old. That is, γ and the ratio HI

HII do not change the quality

of relationship between dual threshold and EUT thresh- old: whether dual threshold will be above or below the EUT threshold depends only on a BI

HI ratio.

It should be noted that the identical derivations can be obtained by applying the concept of expected regret (instead of EUT) [16-19,35]. Although it can be argued that regret is a powerful emotion influencing all cognitive pro- cesses (as so called, “cognitive emotion”) [37,38], and so it may function at level of both system I and system II [39], most authors recognize the affect value of regret [2,10]. Hence, we assumed that regret functions at system I level [2]. Therefore, in our model we restrict the influence of re- gret to system I. Incidentally, our Equation 3) can also be derived from the general Mukherjee’s DSM model even if regret is not specifically invoked [11]. Although Equation 3) implies exact calculations, it

should not be understood as one that provides precise mathematical account of human decision-making. Rather, it should be considered more as a semi-quantitative or qualitative description of the way physicians may make their decisions. First, this is because system I does not per- form exact calculations, but rather relies on “gist” [30,31]

Djulbegovic et al. BMC Medical Informatics and Decision Making 2012, 12:94 Page 5 of 13 http://www.biomedcentral.com/1472-6947/12/94

for assessment of benefits and harms in more qualitative manner. The mechanism depends on associations, emo- tions (so called, “risk as feelings” estimates [10]), as well as memory, and experience [2,5,8,31]. In this sense, the sec- ond part of Equation 3) that relies on system I can be understood as the qualitative modifier (“weight”), which, depending on the system I’s estimates of benefits and harms increases or decreases the first part of equation (which is dependent on system’s II precise usage of evi- dence for benefits and harms). Second, the threshold probability itself should be considered as an “action threshold”- at some point, a physician decides whether to administer treatment or not. Typically, she contrasts the estimated probability of disease against the threshold and acts: if the probability of disease is above the “action threshold”, the physician administers the treatment; if it is below, she decides not to give treatment. So, one way to interpret Equation 3) is to consider physician’s estimate of “gist” of the action threshold: if in her estimation, overall benefits of treatment outweigh harms, and she considers that it is “likely” that the probability of disease is above the threshold probability, then she would act and administer treatment. If the physician assesses that it is “unlikely” that the probability disease is above the “action threshold”, then she would not prescribe the treatment.

The behavior of DSM-M model The exact cognitive mechanisms that underlie dual system processes are not fully elucidated. As discussed through- out this paper, many factors affect dual processes reason- ing leading to suggestions that these processes should be grouped according to the prevailing mechanisms [27]. Fo- cusing on each of these processes may lead to specific the- oretical proposals. Our goal in this paper is to provide overarching cognitive architecture encompassing general features of the majority existing theoretical concepts, while at the same time concentrating on specifics of medical de- cision-making. In general, dual processing theories [27] fall into two main groups [27,40] parallel competitive the- ories and default-interventionalist theories. The parallel- competitive theories assume that system I and II processes proceed in parallel, each competing for control of the re- sponse [27]. If there is a conflict, it is not clear which mechanism is invoked to resolve the conflict [27]. On the other hand, default-interventionist theories postulate that system I generates a rapid and intuitive default response, which may or may not be intervened upon by subsequent slow and deliberative processed of system II [2,5,27]. This can be further operationalized via several general mechan- isms that have been proposed in the literature:

1) Mukherjee’s additive model as described above [11]. It can be categorized as a variant of parallel- competitive theory as it assumes that system I and II

processes proceed in parallel, but does include parameter γ, which can trigger greater or smaller activation of system I. Mukherjee’s model, however, does not explicitly model the choices in terms of categorical decisions (i.e. accept vs. do not accept a given hypothesis), which is a fundamental feature of dual-processing models [27].

2) System I and system II operate on a continuum [41], but in such a way that system I never sleeps [2]. A final decision depends on the activation of both systems I and II [2]. It has been estimated that about 40-50% of decisions are determined by habits (i.e. by system I) [42]. This is also a variation of parallel- competitive theory; it should be noted that latest literature is moving away from this model [5,27].

3) The final decision appears to depend both on the system I and system II in such a way that system I is the first to suggest an answer and system II endorses it [2]. In doing so, system II can exert the full control over system I (such as when it relies on the EUT modeling) or completely fail to oversee functioning of system I (e.g., because of its ignorance or laziness) [2]. Therefore, according to this model, decisions are either made by system I (default) or system II (which may or may not intervene). This is a default- interventionalist model.

4) The variation of the model #3 is the so called “toggle model”, which proposes that decision-maker constantly uses cognitive processes that oscillate between the two systems (toggle) [6,7,9]. This is a variant of default-interventionalist model.

Note that γ is continuous in our model, but it can be made categorical [0,1] if the “toggle” theory is considered to be the correct one. In this case, a logical switch can be introduced in the decision tree to allow toggling between the two systems. Most importantly, by linking Mukherjee’s additive model with the threshold model, we provide the architecture for reconciling parallel competitive theories with default-interventionalist theories. We do it by making explicit that decisions are categorical (via threshold) at certain degree of cognitive effort (modeled via γ) param- eter [27]. That is, the key question is what processes deter- mine acceptance or rejection of a particular (diagnostic) hypothesis. Our model shows that this can occur if we maintain parallel-competing architecture of Mukherjee’s additive model but assume a switch, yes or no answer, whether to accept or reject a given hypothesis (at the threshold). It is evaluation of the (diagnostic) event with respect to the threshold that serves as the final output of our decision-making and reasoning processes. As our model shows, this depends on assumption of parallel working of both system I and system II, and the switch in control of one system over another according to default-

Djulbegovic et al. BMC Medical Informatics and Decision Making 2012, 12:94 Page 6 of 13 http://www.biomedcentral.com/1472-6947/12/94

interventionalist hypothesis. Note that depending on acti- vation of γ parameter and assessment of benefits (gains) and harms (losses) the control can be exerted by either system: sometimes it will be the intuitive system that it will exert the control and our action will take the form “feeling of rightness” [43]; sometimes, it will be system II that it will prevail and drive our decisions. Thus, we succeed in uniting parallel competitive with default- interventionalist models by linking Mukherjee’s additive model with the threshold model for decision-making. As discussed above, many factors can activate the

switch such as the presence or absence of empirical, quantitative data, the context of decision making (e.g. affect poor or rich), the decision maker’s expertise and experience, etc. In addition, extensive psychological re- search has demonstrated that people often use a simple heuristic, which is based on the prominent numbers as powers of 10 (e.g., 1,2,5,10,20,50,100,200 etc.) [44]. That is, although system I does not perform the exact calcula- tions, it still does assess “gist” of relative benefits and harms, and likely does so according to “1/10 aspiration level” [44] (rounded to the closest number) in such a way that the estimates of benefits/harms ratio change by 1,2,5, 10, etc. orders of magnitude. Therefore, in this sec- tion we consider several prototypical situations: 1) when γ = 0, 0.5, or 1; 2) when BII> >HII, BII =HII and BII < < HII; and 3) when regret of omission (BI) < < regret of commission (HI), BI =HI, or BI> >HI

First, note that γ=0, when the numerator of the left frac- tion in the Equation 6 (Additional file 1: Appendix) is zero, i.e., when pBII � 1� pð ÞHII ¼ 0, or solving for p, we obtain p ¼ 1

1þBII HII

, which is exactly the value of the EUT

threshold for the probability at which the expected utilities of the two options are the same. This will correspond to model #3 above, in which system II exerts full control over decision-making. Therefore, when γ = 0, we have the clas- sic EUT and therapeutic threshold model. In this case, re- gret does not affect the EUT benefits and harms, and

pt ¼ HII HIIþBII

¼ 1 1þBII

HII

. If BII> >HII, pt approaches zero and a

decision-maker will recommend treatment to virtually everyone. On the other hand, if BII =HII, pt equals 0.5 and she might recommend treatment if the disease is as likely as not. Finally, if BII < < HII, pt approaches 1.0, and the decision-maker is expected to recommend treatment only if she is absolutely certain in diagnosis. At the other extreme, if γ = 1, we have the pure sys-

tem I model (corresponding to model #3 above, which solely relies on system I processes). Note the value of γ=1, when the denominator of the second fraction in Equation 6 (Additional file 1: Appendix) equals one, or when the expression HI � BI ¼ 0 , i.e., when BI=HI. Under these conditions, it is fairly obvious that the

system I assessments become irrelevant if the perceived net benefit of the treatment is equal to the perceived net harm. When γ=1, regret avoidance becomes the key mo- tivator, not EUT’s benefits and harms. Note that in sys- tem I p is not related to γ in terms of the valuation (Equation 1). Under these circumstances only decision- making under system I operate and the analytical pro- cesses of system II are suppressed (Equation 1) as seen in those decision-makers who tend to follow intuition only, or are extremely affected by their past experiences without considering new facts on the ground. That is, differences in probability do not play any role in such decisions, because a person who only uses system I doesn’t consider probability as a factor. Finally, if γ = 0.5, the decision maker is motivated by

EUT and by regret avoidance (model #2 listed above). In this case, the benefits (BII), harms (HII), regrets of omission (BI) and commission (HI) are all active players. These three cases are presented in Table 1 (see Additional file 2) which shows threshold probabilities for γ=0.5 and objective data indicating a high benefit/harms ratio (BII=HII ¼ 10). Also shown is how the threshold probability depends on indi- vidual risk perception. If HI> >HI, it magnifies effect of BI/HI (see Equation 3), which results in extreme behavior in sense of increasing likelihood that such a person will ei- ther always accept (as pt<0) or reject treatment (as pt>1). For HI <<HII, the impact on the way system I processes benefits and harms is not that pronounced and influences the EUT threshold to much smaller extent.

Results Illustrative medical examples Clinical examples abound to illustrate applicability of our model. To illustrate the salient points of our model, we chose two prototypical examples where there is close trade-offs between treatments’ benefits and harms.

Example #1: treatment of pulmonary embolism Pulmonary embolism (PE) (blood clot in the lungs) is an important clinical problem that can lead to significant morbidity and death [45]. Even though many diagnostic imaging tests exist to aid in the accurate diagnosis of PE, the tests are often inconclusive, and physicians are left to face the decision whether to treat patient for presumptive PE, or attribute the patient’s clinical presentation (such as shortness of breath and/or chest pain) to other possible etiologies. There exists an effective treatment for a PE, which consists of the administration of 2 anticoagulants (blood thinners): heparin followed by oral anticoagulants such as warfarin [46,47]. Heparin (unfractionated or low- molecular weight heparins) are highly effective treatments associated with relative risk reduction of death from PE by 70-90% in comparison to no treatment [46,47]. This con- verts into the absolute death reduction as: net benefits,

Djulbegovic et al. BMC Medical Informatics and Decision Making 2012, 12:94 Page 7 of 13 http://www.biomedcentral.com/1472-6947/12/94

BII=17.5% to 22.5% (calculated as 25% morality without heparin minus 7.5% to 2.5% with heparin) [17,18,46,47]. However, these drugs are also associated with a significant risk of life-threatening bleeding; net harms range from HII=0.037% (a typical scenario) to 5% (a worst-case sce- nario) depending on the patients’ other comorbid condi- tions [17,18,47,48]. Thus, net benefits/net harms range from 60.8 (22.5/0.037) (best case) to 3.5 (17.5/5)(worst case scenario). If we apply a classic EUT threshold [13,14,36], which relies solely on system II processes, we observe that the probability of pulmonary embolism above which the physician should administer anticoagulants ranges from 1.6% [¼ 1= 1þ 60:8ð Þ ] (best case) to 22.2% [1= 1þ 3:5ð Þ ](worst case scenario). However, ample clin- ical experience has demonstrated that few clinicians would consider prescribing anticoagulants at such low probability of PE [18]. In fact, most experts in the field recommend giving anticoagulants when probability of PE exceeds 95% [49-51]. We have previously suggested that this is because regret associated with administering unnecessary and po- tentially harmful treatments under these circumstances likely outweighs regret associated with failing to adminis- ter potentially beneficial anticoagulants [17-19]. We now show how this argument can be made in the context of dual processing theory. Indeed, some physicians may feel that the risk of bleeding may be much higher, particularly in case of a patient who recently experienced major hemorrhage. The physician may not have data readily available to adjust her EUT, system II-based calculations. Rather, she employs the system I-based reasoning, globally assessing the benefits and harms of treatments under her disposal. Importantly, these are personal, intuitive, affect- based, subjective judgments of the values of outcomes that are influenced by memory limitations and recent experi- ences and that may not be objectively based on the exter- nal evidence [2,30-33]. In addition, it is well documented that the physicians’ recent experience leads to a type of bias, known as primacy effect, that is governed by system I [2,33]. If the last patient with PE whom the physician took care of had severe bleeding, system I may be primed in such a way that it will likely conclude that harms outweigh benefits. In our case of PE, if her reasoning is dominated by system I (operating, say, at γ level of 0.77 according to model #2 listed above, see Section “The behavior of DSM- M model”) in a such way that the physician concludes that if harms is larger than benefits by 10%, then the threshold probability above which she will treat her patient sus- pected of PE exceeds 95% [as easily demonstrated after plugging in the benefits/harms values in Equation 3) pt dualð Þ ¼ :222� :77ð Þ= 2�:23ð Þ� �:10=:225ð Þ ¼ 0:966 ¼ 96:6% for k ¼ 1]. Note that this calculation describes cir- cumstances under which the physician would adhere to the contemporary practice guidelines i.e. to prescribe anticoagulants when PE exceeds 95% [49-51]. It should be

further noted that if γ value is only slightly higher (≥0.78), the physician will require the absolute certainty to act (i.e. the threshold ≥1). DSM offers an account of the opposite behavior as well

i.e. the threshold based on global evaluation using both system I and system II can also be lower than the EUT threshold (if BI>HI additive, model #1, Equation 3). For example, the physician may trivialize the risks of treatment and believe that the benefits are much higher than the treatment harms. As a result, the threshold above which she commits to treatment drops below EUT threshold (as predicted by Equation 3). Figure 3 shows how the decision threshold (pt) is affected by the relative involvement of systems I and II in dual process model of medical decision-making in the “best” ( BII=HII ¼ 60:8 ) and “worst case” scenario (BII=HII ¼ 3:5) for treatment of PE and when system I valuation of benefits is greater than harms or when harms are perceived to outweigh benefits. It can be seen that when objective data indicate that bene- fits considerably outweigh harms (BII >>> HII) (as when BII=HII ¼ 60:8), then as long as system I values benefits as being greater than harms, the threshold dramatically drops to zero indicating that the extent of system I involvement (i.e. γ value) in decision-making is of little consequence. However, if system I clashes with objective data, then the probability of PE above which the decision-maker is pre- pared to treat, dramatically increases (Figure 3a). Similarly, in all other circumstances (when BII >HII, BII ~HII, BII < HII), the threshold probability is significantly affected by involvement of system I (Figures 3b–3d).

Example #2: treatment of acute leukemia Acute myeloid leukemia (AML) is a life-threatening dis- ease, which, depending on the aggressiveness of disease can be cured in the substantial minority of patients. To achieve a cure, patients are typically given induction chemotherapy to bring the disease into remission, after which another form of intensive therapy – so called, con- solidation treatment – is given. To achieve a cure in patients with more aggressive course of disease such as those classified as intermediate- and poor-risk AML based on cytogenetic features of disease, allogeneic stem cell transplant (alloSCT) is recommended [52]. However, the cure is not without price- many patients given alloSCT as a consolidation therapy die due to treatment. A decision dilemma faced by a physician is whether to recommend alloSCT, or alternative treatment, such as chemotherapy or autologous SCT, which has lower cure rate but less treatment-related mortality. In intermediate-risk AML, for example, credible evidence shows that, compared with chemotherapy allogeneic alloSCT result in better leukemia-free survival (LFS) by at least 12% at 4 years (LFS with alloSCT =53% vs 41% with chemotherapy/auto SCT) [53]. Treatment-related

a) BII >>> HII b) BII > HII

d) BII > HIIc) BII= HII

Figure 3 Dual decision threshold (pt) as the function of relative involvement of systems I and II in dual process model of medical decision making: a) objective data show very high benefit/harms ratio (BII> >HII), b) moderately high benefit/harms ratio (BII > HII), c) BII = HII, d) BII < HII. The intercept at y axis threshold probability. The graph shows how the threshold is affected by the extent of system I involvement (γ) and whether system I perceives that benefits is greater than harms [by 5% in this example](red lines, circles) or that harms outweigh benefits[by 5%] (blue line, squares). Decision-maker accepts treatment if the probability of disease exceeds the threshold; otherwise, treatment would not be acceptable (see text for details).

Djulbegovic et al. BMC Medical Informatics and Decision Making 2012, 12:94 Page 8 of 13 http://www.biomedcentral.com/1472-6947/12/94

mortality is much higher with alloSCT by 16%, on average (19% with alloSCT vs. 3% with chemotherapy/ autoSCT) [53]. This means that based on objective data, and using rational EUT model, we should recom- mend alloSCT for any probability of AML relapse ≥57.1% threshold ¼ 1= 1þ 0_12=0:16ð Þ ¼ 0:571ð Þ. There- fore, treatment benefits and harms are, on average, very close. Because of this, the driving force to recommend alloSCT is the physician’s estimates of the patient’s toler- ability of alloSCT: if she assess that the patient will not be able to tolerate alloSCT, the physician will not recom- mend transplant. Conversely, if she thinks that the pa- tient will be able to tolerate allo SCT, the physician will recommend it. Although there are objective criteria to evaluate a patient’s eligibility for transplant, the assess- ment to the large extent depends on physicians’ judg- ment and experience [54]. That is, the assessment of patient’s eligibility for transplant depends both on the

objective data on benefits and harms (system II ingredi- ents) and intuitive, gist type of judgment (characteristics of system I). As discussed above, system I does not con- duct the precise calculations. Rather, it relies on “gist” or on simple heuristics such as those that are based on powers of 10 (e.g., 1,2,5,10,20, etc.) [42]. The physician, therefore, adjusts the threshold above or below based on her intuitive calculations. For instance, it is often the case that the physician whose patient recently died dur- ing the transplant is more reluctant to recommend the procedure even to those patients who, otherwise, seems fit for it. In doing so, the physician in fact modifies her/ his dual system threshold upwards. In our example, let’s assume that the physician judges that the harms of alloSCT for a given patient is twice as large as reported in the studies where patients were carefully selected for transplant [52]. That, in our case, would mean that mor- tality due to alloSCT is 32% (instead of 16%). We can

Djulbegovic et al. BMC Medical Informatics and Decision Making 2012, 12:94 Page 9 of 13 http://www.biomedcentral.com/1472-6947/12/94

now plug these numbers in Equation 3) (BII = 0.12, HII = 0.16, BI = 0.12, HI = 0.32). Note that the physician can make this judgment at vari-

ous level of activation of system I. If the decision is pre- dominantly driven by system I judgment then our physician’s threshold according to Equation 3) is greater than 100% for all circumstances in which γ value exceeds 55%. That means that under these circumstances of sys- tem I activation, the physician will never recommend transplant. The opposite can occur for those physicians whose experience is not affected by poor patients’ out- comes. Under such circumstances, the physician may judge the patient to be in such a good condition that she may re-adjust the reported treatment-related transplant risk to be as half of those observed risks in the published clinical studies (i.e. 8%). The new numbers required to de- termine the threshold according to Equation 3 are: BII = 0.12, HII = 0.16, BI = 0.12, HI = 0.08. If the physician relies excessively on system I, as often seen in busy clinics where decisions are routinely made on “automatic pilot”, the dual threshold drops to zero (for all γ >89%). That means, that the physician will recommend alloSCT to all her/his patients under these circumstances. As discussed above, we provide the precise calculations

only to illustrate the logic of decision-making. The process should be understood more along semi-quantitative or qualitative description of clinical decision-making. Although currently the Equation 3) allows entry of almost any value for benefit and harms, it is probably the case that benefit and harms as perceived by system I are based on “1/10 aspirational level” [44], so that only values of 1,2,5,10, 20 etc. should be allowed. This is, however, empirical question that should be answered in further experimental testing; there- fore, at this time, we decided not to provide the exact boundaries of the values for benefit and harms that can be entered in Equation 3 (see Discussion). Note also that these calculations are decision-maker specific, and although we il- lustrate them from the perspective of the physician, the same approach applies to the patient, who ultimately has to agree –based on her own dual cognitive processing- on the suggested course of treatment actions.

Discussion Models of medical decision-making belong to two gen- eral classes-descriptive and prescriptive. The former, which the DSM-M exemplifies, attempt to explain why decision makers take or might take certain actions when presented with challenging decision problems abundant in contemporary medicine. The latter, exemplified by the normative therapeutic threshold models [13,14] pre- scribe the choice options that a rational decision maker should take. We have defined the first formal dual- process theory of medical decision-making by taking into consideration the deliberative and the experiential

aspects that encompass many of the critical decisions physicians face in practice. Mathematically, our model represents an extension of Mukharjee’s additive Dual System Model [11] to the clinical situation where a physician faces frequent dilemmas: whether to treat the patient who may or may not have the disease, or choose one treatment over another for prevention of disease that is yet to occur. Our model is unique in that incor- porates an exemplar of strong emotion, decision regret, as one of the important components of system I func- tioning. We focused on regret because previous research has shown that people often violate EUT prescribed choice options in an effort to minimize anticipated re- gret [1,2,20]. Although we use the more common psy- chological term “regret,” the concept is analogous to Feinstein’s term “chagrin” [55]. In fact, explicit consider- ation of post-choice regret in decision making has been considered an essential element in any serious theory of choice and certainly dominates many clinical decisions [1,2,20]. We also reformulated the original model using the threshold concept- a fundamental approach in med- ical decision-making [13,14,36]. The threshold concept represents a linchpin between evidence (which presents on the continuum of credibility) and decision-making, which is a categorical exercise (as choice options are ei- ther selected or not) [13,14,36]. Using an example such as pulmonary embolism, we have shown how the extended model can explain deviations from outcomes predicted by EUT, and account for the variation in man- agement of pulmonary embolism [45]. In general, it is possible that the huge practice variation well documen- ted in contemporary medicine [56-61], can be, in part, due to individual differences in subjective judgments of disease prevalence and “thresholds” at which physicians act. [17,18,62]. This may be because quantitative inter- pretations of qualitative descriptors such as rarely, un- likely, possible, or likely [63] differ markedly among individuals and hence “gist” representations of a given clinical situation can vary widely among different physi- cians [30]. We are, of course, aware that many other fac- tors contribute to variation in patient care including the structure of local care organizations, the availability of medical technologies, financial incentives etc [60]. Our intent in this article is to highlight, yet another import- ant factor- individual differences in risk assessment as shaped by different mechanisms operating within a dual process model of human cognitive functioning [5]. There are many theories of decision-making [64].

Most assume a single system of human reasoning [11]. Nevertheless, all major theories of choice agree that ra- tional decision-making requires integrations of benefits (gains) and harms (losses). EUT vs. non-EUT theories of decision-making differ in how benefits and harms should be integrated in a given decision task. To date, dual

Djulbegovic et al. BMC Medical Informatics and Decision Making 2012, 12:94 Page 10 of 13 http://www.biomedcentral.com/1472-6947/12/94

processing theory provides the most compelling explan- ation how both intuitive and rational cognitive processes integrate information on benefits and harms and provide not only descriptive assessments of decision-making, but possibly may lead to insights that improve the way deci- sions are made. Figures 3 & 4 illustrate how dual decision threshold (shown on the Y axis) for deciding between two possible courses of action can be influenced by the degree of system I involvement. As discussed above and mathem- atically captured in Equation 3, the clinical action such as treat versus no treat is best explained by relating benefit and harms of proposed therapeutic interventions to the threshold probability: if the estimated probability of dis- ease is greater than the threshold probability, then the decision-maker is inclined to give treatment; if the prob- ability of disease is below the threshold, then the treat- ment is withheld. Figure 4 shows a dramatic drop in the decision threshold as a function of the ratio between bene- fit and harms, which is derived from empirically obtained evidence. When these data are solely relied on by system II, the rational course of action consists of administering treatment as long as the probability of disease is above the threshold regardless how low the threshold probability drops [13,14,36] (which in case of the treatment of a pa- tient with pulmonary embolism can be as low as 1.6%!). Paradoxically, if we were to adopt this – presumably most rational-approach to the practice of medicine, we would likely see a further explosion of inappropriate and wasteful use of health care resources [18,21]. This is because in today’s practice, benefits of approved treatments vastly outweigh their harms, and as a result threshold probability values is predictably very low for the majority of health care interventions employed in the contemporary clinical practice [18]. System I, however, does offer a means of mitigation. The correction of the thresholds – our action

Figure 4 Dual Decision Threshold Model. Classic, expected utility thresh system II, EUT (expected utility threshold) (solid line). The treatment should should be withheld. Note that if system I perceives that harms are higher t higher than classic EUT (dotted line). However, if BI > HI, the threshold prob for details).

whether we are comfortable treating at higher or lower probability than the thresholds obtained via usage of sys- tem II – depends on the extent of involvement of system I in decision-making. If system I perceives that harms are higher than system I benefits, the threshold probability is always higher than classic EUT threshold. However, if BI>HI, the threshold probability is always lower than the EUT threshold (Figure 4). This is particularly evident in clinical practice when physicians attempt to tailor evi- dence based on the results of the research study, which generates the “group averages”, to individual patients who often differ in important ways from patients enrolled in the research studies (e.g., these patients may be older, have comorbid conditions, might be using multiple medica- tions, etc.) [65]. It is under these circumstances that sys- tem I affects our judgments and can give rise to different decisions from those based solely on system II. Note, how- ever, that although system I does assess benefits and harms, it likely does so via”gist” representation and not ne- cessarily by employing the exact numerical values as sys- tem II does [30]. System I is also affected by emotions, as illustrated in the case where experts panels of the govern- ments of many countries recommended H1N1 influenza vaccination, but where inoculation was refused by the ma- jority of patients [66,67]. It is interesting to examine circumstances under which

we always treat (pt≤ 0) or never treats (pt≥ 1). Equation 1 (Additional file 2: Table S1) shows that when objective evidence indicates that benefits outweigh harms, and when this is further augmented by the decision-maker’s risk attitude in such a way that it magnifies system I’s valuation of benefits and harms, then we can expect to continue to witness further overtreatment in clinical practice (as pt drops to zero) [65]. However, when the decision-maker perceives the benefits smaller than

old probability as a function of benefit/harms ratio as derived by be given if the probability of disease is above the threshold, otherwise han system I benefits (BI < HI), the threshold probability is always ability is always lower than the EUT threshold (dashed line) (see text

Djulbegovic et al. BMC Medical Informatics and Decision Making 2012, 12:94 Page 11 of 13 http://www.biomedcentral.com/1472-6947/12/94

harms, then the threshold increases; consequently, the decision-maker will require higher diagnostic certainty before acting (Figure 3 & Figure 4). This may occur dur- ing extrapolation of research results from the group averages to individual patients, when empirical evidence about BII and HII is considered to be unreliable, when the decision-maker is risk averse, or when his or her cognitive processes are biased through the distorting effects of recent experience, memory limitations or other forms of biases well described in the literature [2,31,33]. This discussion illustrates how the “rationality of action” may require a re-definition, one encompassing both the formal principles of probability theory and human intui- tions about good decisions [5,68]. Our goal here is not to demonstrate that one approach is conclusively super- ior to the other- we are merely outlining the differences in the current physicians’ behavior from the perspective of dual processing theory. Despite the growing recognition of the importance of

dual processing for decision-making [2,5], a few formal models have been developed to try to capture the es- sence of the way we make decisions. Because different authors focus on different aspects of a multitude of decision-making processes, Evans has recently pointed out that there are many dual processing theories [27] which fall into two main groups [27,40] parallel competi- tive theories and default-interventionalist theories. While the exact accounts of cognitive processes between these two groups of theories differ [27], as discussed above (Section The behavior of DSM-M Model), we, for the first, time provide a platform, albeit the theoretical one, for reconciling parallel competitive theories with default- interventionalist theories. Nevertheless, our main goal is to define a theoretical

model for medical decision-making; such a model may enable creation of new theoretical frameworks for future empirical research. Future research, obviously, involves extension of the model described herein to more com- plex clinical situations beyond relatively simple two- alternative situation, even if the latter is frequently encountered in practice. Particularly interesting will be the extension of our dual processing model to include the use of diagnostic tests as the number of new diag- nostic technologies continues to explode. Finally, and most importantly, the model presented here needs em- pirical verification. This limitation is not unique to our model, however, and this criticism can be leveled against most current medical decision-making models, which are rarely, if ever, subjected to empirical verification. Our model heavily relies on Mukherjee’s model [11],

and is accurate to the extent his additive dual processing model is correct (Figure 1, Equations 1 & 2). Also, note that we have extended Mukherjee’s DSM model by omit- ting his scaling constant k and using general utility

expressions, rather than a single parameter monotonic power function. As discussed above, many factors can ac- tivate the switch of system II. In fact, Kahneman warns [2] that “because you have little direct knowledge what goes on in your mind, you will never know that you might have made a different judgment or reached a different decision under very slightly different circumstances”. This implies that the multiple factors affecting the gamma parameter cannot be directly modeled. A possible solution –and area for future research building on the psychological “fuzzy trace theory” [30]-would be to employ a fuzzy logic model to assess the values of γ (and threshold) as a function of multiple fuzzy inputs [69]. The complexity described here notwithstanding, we

believe that the empirical verification of our current dual processing model is feasible. Even without direct model- ing of all factors affecting γ parameter, our model gener- ates empirically falsifiable qualitative predictions as it clearly identifies circumstances under which the decision threshold is increased or decreased as a function of acti- vation of system I (γ parameter). Using simulation to imitate the various real-life decision-making scenarios [70] offers most logical avenue toward the first empirical testing of our model. Our model also holds promise in medical education.

As highlighted in Introduction, modern knowledge of cognition has taught us that most people, including phy- sicians process information using both system I (fast, in- tuitive) and system II (slow, deliberative) reasoning at different times but few investigators have examined how to teach physicians to integrate both modes of reasoning in arriving at therapeutic strategies. On the diagnostic side, many investigators [6,71] have examined clinical reasoning and proposed how experienced physicians move between system I and system II, although most early papers used different terminology. The integration of system I and system II in therapeutic decision making in medicine has been less well examined. A number of investigators have proposed approaches to using and teaching system II reasoning, including the use of deci- sion models [71]. Although this is taught in some schools it has not yet taken medical education by storm [71]. In the field of economic analysis Mukerjee has proposed a theoretical means of combining system I and system II reasoning. In this paper, we build on Mukurjee’s work and show how the integration of system I and sys- tem II therapeutic reasoning can form a basis for teaching students and experienced physicians to recognize and integrate system I and system II reasoning. Our model uniquely captures most salient features of (medical) decision-making, which can be effectively employed for di- dactic purposes. It is believed that by recognizing separate roles of system II and the influence of system I mechan- isms on the way we make decisions, we can be in a better

Djulbegovic et al. BMC Medical Informatics and Decision Making 2012, 12:94 Page 12 of 13 http://www.biomedcentral.com/1472-6947/12/94

position to harness both types of processes toward better practice of making clinical decisions [2,9].

Conclusion We hope that our model will stimulate new lines of em- pirical and theoretical work in medical decision-making. In summary, we have described the first dual processing model of medical decision-making, which has potential to enrich the current medical decision-making field dominated by expected utility theory.

Additional files

Additional file 1: Appendix: Derivation of DSM-M equation.

Additional file 2: Table S1. Evaluation of Behavior of Dual Processing Model for Medical Decision-Making (DSM -M). Threshold probability as a function of individual risk perception.

Competing interests The authors declare that they have no competing interests.

Authors' contributions BD had an idea for the study. BD & IH jointly developed the model. IH solved the model. BD and IH performed the analyses. JB and AT performed additional analyses. SGP analyzed the performance of dual processing model and provided an additional intellectual input. BD wrote the first draft. All authors read and approved the final manuscript.

Acknowledgments We want to thank to Dr Shira Elqayam of De Montfort University, Leicester, UK for the most helpful comments, in particular to introducing us to a notion of the parallel competitive vs. default-interventionalist dual processing theories and pointing the way how our model can help reconcile these two competing theoretical frameworks. Presented as a poster at: 14th Biennial European Conference of the Society for Medical Decision Making (SMDM Europe 2012) Oslo, Norway, June 10–12, 2012. Supported by the US DoA grant #W81 XWH 09-2-0175 (PI Djulbegovic).

Author details 1Center for Evidence-based Medicine and Health Outcomes Research, Tampa, FL, USA. 2Department of Internal Medicine, Division of Evidence- based Medicine and Health Outcomes Research University of South Florida, Tampa, FL, USA. 3Departments of Hematology and Health Outcomes and Behavior, H. Lee Moffitt Cancer Center & Research Institute, Tampa, FL, USA. 4Department of Mathematics, Indiana University Northwest, Gary, IN, USA. 5College of Nursing, University of South Florida, Tampa, FL, USA. 6Division of Clinical Decision Making, Department of Medicine, Tufts Medical Center, Boston, USA. 7USF Health, 12901 Bruce B. Downs Boulevard, MDC27, Tampa, FL 33612, USA.

Received: 18 June 2012 Accepted: 21 August 2012 Published: 3 September 2012

References 1. Kahneman D: Maps of bounded rationality: psychology for behavioral

economics. Am Econ Rev 2003, 93:1449–1475. 2. Kahnemen D: Thinking fast and slow. New York: Farrar, Straus and Giroux; 2011. 3. Evans JSTBT: Hypothethical thinking. Dual processes in reasoning and

judgement. New York: Psychology Press: Taylor and Francis Group; 2007. 4. Stanovich KE, West RF: Individual differences in reasoning:implications for

the rationality debate? Behavioral and Brain Sciences 2000, 23:645–726. 5. Stanovich KE: Rationality and the Reflective Mind. Oxford: Oxford University

Press; 2011. 6. Croskerry P: Clinical cognition and diagnostic error: applications of a dual

process model of reasoning. Adv Health Sci Educ Theory Pract 2009, 14(Suppl 1):27–35.

7. Croskerry P: A universal model of diagnostic reasoning. Acad Med 2009, 84(8):1022–1028.

8. Croskerry P, Abbass A, Wu AW: Emotional influences in patient safety. J Patient Saf 2010, 6(4):199–205.

9. Croskerry P, Nimmo GR: Better clinical decision making and reducing diagnostic error. J R Coll Physicians Edinb 2011, 41(2):155–162.

10. Slovic P, Finucane ML, Peters E, MacGregor DG: Risk as analysis and risk as feelings: some thoughts about affect, reason, risk, and rationality. Risk Anal 2004, 24(2):311–322.

11. Mukherjee K: A dual system model of preferences under risk. Psychol Rev 2010, 177(1):243–255.

12. Djulbegovic B, Hozo I, Lyman GH: Linking evidence-based medicine therapeutic summary measures to clinical decision analysis. MedGenMed 2000, 2(1):E6.

13. Pauker S, Kassirer J: Therapeutic decision making: a cost benefit analysis. N Engl J Med 1975, 293:229–234.

14. Pauker SG, Kassirer J: The threshold approach to clinical decision making. N Engl J Med 1980, 302:1109–1117.

15. Djulbegovic B, Desoky AH: Equation and nomogram for calculation of testing and treatment thresholds. Med Decis Making 1996, 16(2):198–199.

16. Djulbegovic B, Hozo I, Schwartz A, McMasters K: Acceptable regret in medical decision making. Med Hypotheses 1999, 53:253–259.

17. Hozo I, Djulbegovic B:When is diagnostic testing inappropriate or irrational? Acceptable regret approach. Med Decis Making 2008, 28(4):540–553.

18. Hozo I, Djulbegovic B: Will insistence on practicing medicine according to expected utility theory lead to an increase in diagnostic testing? Medical Decision Making 2009, 29:320–322.

19. Hozo I, Djulbegovic B: Clarification and corrections of acceptable regret model. Medical Decision Making 2009, 29:323–324.

20. Kahneman D, Wakker PP, Sarin RK: Back to Bentham? Explorations of experienced utility. Q J Econ 1997, 112:375–405.

21. Berwick DM, Hackbarth AD: Eliminating waste in US health care. JAMA: The Journal of the American Medical Association 2012, 307(14):1513–1516.

22. Manchikanti L, Falco FJ, Boswell MV, Hirsch JA: Facts, fallacies, and politics of comparative effectiveness research: part 2 – implications for interventional pain management. Pain Physician 2010, 13(1):E55–E79.

23. Manchikanti L, Falco FJ, Boswell MV, Hirsch JA: Facts, fallacies, and politics of comparative effectiveness research: part I. Basic considerations. Pain Physician 2010, 13(1):E23–E54.

24. Hsee CK, Rottenstreich Y: Music, pandas and muggers: on the affective psychology of value. J Exp Psychol 2004, 133:23–30.

25. Rottenstreich Y, Hsee CK: Money, kisses, and electric shock: On the affective psychology of risk. Psychol Sci 2001, 12:185–190.

26. Evans JSTBT: Thinking Twice. Two Minds in One Brain. Oxford: Oxford University Press; 2010.

27. Evans JSTBT: Dual-process theories of reasoning: contemporary issues and developmental applications. Dev Rev 2011, 31:86–102.

28. Edwards W, Miles R Jr, vonWinterfeld D: Advances in decision analysis. From foundations to applications. New York: Cambridge University Press; 2007.

29. Simon HA: Information processsing models of cognition. Ann Review Psychol 1979, 30:263–296.

30. Reyna VF, Brainerd CJ: Dual processes in decision making and developmental neuroscience: a fuzzy-trace model. Dev Rev 2011, 31(2–3):180–206.

31. Reyna VF, Hamilton AJ: The importance of memory in informed consent for surgical risk. Med Decis Making 2001, 21(2):152–155.

32. Kahneman D, Tversky A: The psychology of preferences. Sci American 1982, 246:160–173.

33. Tversky A, Kahneman D: Judgements under uncertainty: heuristics and biases. Science 1974, 185:1124–1131.

34. Djulbegovic B, Hozo II, Fields KK, Sullivan D: High-dose chemotherapy in the adjuvant treatment of breast cancer: benefit/risk analysis. Cancer Control 1998, 5(5):394–405.

35. Djulbegovic B, Hozo I: When should potentially false research findings be considered acceptable? PLoS Medicine 2007, 4(2):e26.

36. Djulbegovic B, Hozo I: Linking evidence-based medicine to clinical decision analysis. Med Decision Making 1998, 18:464. abstract.

37. Zeelenberg M, Pieters R: A theory of regret regulation 1.0. J Consumer Psychol 2007, 17:3–18.

38. Zeelenberg M, Pieters R: A theory of regret regulation 1.1. J Consumer Psychol 2007, 17:29–35.

Djulbegovic et al. BMC Medical Informatics and Decision Making 2012, 12:94 Page 13 of 13 http://www.biomedcentral.com/1472-6947/12/94

39. Tsalatsanis A, Hozo I, Vickers A, Djulbegovic B: A regret theory approach to decision curve analysis: a novel method for eliciting decision makers’ preferences and decision-making. BMC Medical Informatics and Decision Making 2010, 10(1):51.

40. Evans JSTBT: On the resolution of conflict in dual process theories of reasoning. Think Reasoning 2007, 13(4):321–339.

41. Hammond KR: Human judgment and social policy. irreducible uncertainty, inevitable error, unavoidable injustice. Oxford: Oxford University Press; 1996.

42. Duhigg C: The power of habit: why we do what we do in life and business. New York: Random House; 2012.

43. Thompson VA, Prowse Turner JA, Pennycook G: Intuition, reason, and metacognition. Cogn Psychol 2011, 63:107–140.

44. Brandstatter E, Gigerenzer G: The priority heuristic: making choices without trade-offs. Psychol Rev 2006, 113:409–432.

45. Sox HC: Better care for patients with suspected pulmonary embolism. Ann Intern Med 2006, 144(3):210–212.

46. Barritt DW, Jordan SC: Anticoagulant drugs in the treatment of pulmonary embolism. A controlled trial. Lancet 1960, 1:1309–1312.

47. Segal JB, Eng J, Jenckes MW, Tamariz LJ, Bolger DT, Krishnan JA, Streiff MB, Harris KA, Feuerstein CJ, Bass EB: Diagnosis and treatment of deep venous thrombosis and pulmonary embolism. In AHRQ Publication No 03-E016. Washington, DC: Agency for Healthcare Research and Quality, U.S. Department of Health and Human Services; 2003.

48. Linkins L-A, Choi PT, Douketis JD: Clinical impact of bleeding in patients taking oral anticoagulant therapy for venous thromboembolism: a meta-analysis. Ann Intern Med 2003, 139(11):893–900.

49. Roy PM, Durieux P, Gillaizeau F, Legall C, Armand-Perroux A, Martino L, Hachelaf M, Dubart AE, Schmidt J, Cristiano M, et al: A computerized handheld decision-support system to improve pulmonary embolism diagnosis: a randomized trial. Ann Intern Med 2009, 151(10):677–686.

50. Roy P-M, Colombet I, Durieux P, Chatellier G, Sors H, Meyer G: Systematic review and meta-analysis of strategies for the diagnosis of suspected pulmonary embolism. BMJ 2005, 331(7511):259.

51. Hull RD: Diagnosing pulmonary embolism with improved certainty and simplicity. JAMA 2006, 295(2):213–215.

52. Koreth J, Schlenk R, Kopecky KJ, Honda S, Sierra J, Djulbegovic BJ, Wadleigh M, DeAngelo DJ, Stone RM, Sakamaki H, et al: Allogeneic stem cell transplantation for acute myeloid leukemia in first complete remission: systematic review and meta-analysis of prospective clinical trials. Jama 2009, 301(22):2349–2361.

53. Cornelissen JJ, Van Putten WL, Verdonck LF, Theobald M, Jacky E, Daenen SM, van Marwijk Kooy M, Wijermans P, Schouten H, Huijgens PC, et al: Results of a HOVON/SAKK donor versus no-donor analysis of myeloablative HLA-identical sibling stem cell transplantation in first remission acute myeloid leukemia in young and middle-aged adults: benefits for whom? Blood 2007, 109(9):3658–3666.

54. Djulbegovic B: Principles of reasoning and decision-making. In Decision Making in Oncology Evidence-based management. Edited by Djulbegovic B, Sullivan DS. New York: Churchill Livingstone, Inc; 1997:1–14.

55. Feinstein AR: The ‘chagrin factor’ and qualitative decision analysis. Arch Intern Med 1985, 145:1257–1259.

56. Detsky AS: Regional variation in medical care. N Engl J Med 1995, 333:5890590.

57. Dilts DM: Practice variation: the Achilles’ Heel in quality cancer care. J Clin Oncol 2005, 23(25):5881–5882.

58. Eddy DM: Variations in physician practice: the role of uncertainty. Health Aff 1984, 3(2):74–89.

59. Fisher ES, Wennberg DE, Stukel TA, Gottlieb DJ, Lucas FL, Pinder EL: The implications of regional variations in medicare spending. Part 2: health outcomes and satisfaction with care. Ann Intern Med 2003, 138(4):288–298.

60. Sirovich BE, Gottlieb DJ, Welch HG, Fisher ES: Regional variations in health care intensity and physician perceptions of quality of care. Ann Intern Med 2006, 144(9):641–649.

61. Zhang Y, Baicker K, Newhouse JP: Geographic variation in medicare drug spending. N Engl J Med 2010, 363(5):405–409.

62. Hozo I, Djulbegovic B: Explaining variation in practice: acceptable regret approoach. Boston: 28th Annual Meeting of the Society for Medical Decision Making; 2006.

63. Shaw NJ, Dear PR: How do parents of babies interpret qualitative expressions of probability? Arch Dis Child 1990, 65(5):520–523.

64. Reyna VF: Theories of medical decision making and health: an evidence- based approach. Med Decis Making 2008, 28(6):829–833.

65. Djulbegovic B, Paul A: From efficacy to effectiveness in the face of uncertainty: indication creep and prevention creep. JAMA 2011, 305 (19):2005–2006.

66. Ofri D: The emotional epidemiology of H1N1 influenza vaccination. N Engl J Med 2009, 361(27):2594–2595.

67. Poland GA: The 2009–2010 influenza pandemic: effects on pandemic and seasonal vaccine uptake and lessons learned for seasonal vaccination campaigns. Vaccine 2010, 28(Supplement 4(0)):D3–D13.

68. Krantz DH, Kunreuther HC: Goals and plans in decision making. Judgement and Decision Making 2007, 2(3):137–168.

69. Zimmerman HJ: Fuzzy set theory anf its applications. 3rd edition. Boston: Kluwer; 1996.

70. Society for simulation in healthcare. http://ssih.org/about-simulation (Last accessed: August 27,2012).

71. Kassirer JP, Kopelman RI: Learning clinical reasoning. Baltimore: Williams & Wilkins; 1991.

doi:10.1186/1472-6947-12-94 Cite this article as: Djulbegovic et al.: Dual processing model of medical decision-making. BMC Medical Informatics and Decision Making 2012 12:94.

Submit your next manuscript to BioMed Central and take full advantage of:

• Convenient online submission

• Thorough peer review

• No space constraints or color figure charges

• Immediate publication on acceptance

• Inclusion in PubMed, CAS, Scopus and Google Scholar

• Research which is freely available for redistribution

Submit your manuscript at www.biomedcentral.com/submit

<< /ASCII85EncodePages false /AllowTransparency false /AutoPositionEPSFiles true /AutoRotatePages /PageByPage /Binding /Left /CalGrayProfile (Dot Gain 20%) /CalRGBProfile (sRGB IEC61966-2.1) /CalCMYKProfile (U.S. Web Coated 50SWOP51 v2) /sRGBProfile (sRGB IEC61966-2.1) /CannotEmbedFontPolicy /Error /CompatibilityLevel 1.4 /CompressObjects /Tags /CompressPages true /ConvertImagesToIndexed true /PassThroughJPEGImages true /CreateJobTicket false /DefaultRenderingIntent /Default /DetectBlends true /DetectCurves 0.0000 /ColorConversionStrategy /LeaveColorUnchanged /DoThumbnails true /EmbedAllFonts true /EmbedOpenType false /ParseICCProfilesInComments true /EmbedJobOptions true /DSCReportingLevel 0 /EmitDSCWarnings false /EndPage -1 /ImageMemory 1048576 /LockDistillerParams true /MaxSubsetPct 100 /Optimize true /OPM 1 /ParseDSCComments true /ParseDSCCommentsForDocInfo true /PreserveCopyPage true /PreserveDICMYKValues true /PreserveEPSInfo true /PreserveFlatness true /PreserveHalftoneInfo false /PreserveOPIComments false /PreserveOverprintSettings true /StartPage 1 /SubsetFonts true /TransferFunctionInfo /Apply /UCRandBGInfo /Preserve /UsePrologue false /ColorSettingsFile () /AlwaysEmbed [ true ] /NeverEmbed [ true ] /AntiAliasColorImages false /CropColorImages true /ColorImageMinResolution 300 /ColorImageMinResolutionPolicy /OK /DownsampleColorImages true /ColorImageDownsampleType /Bicubic /ColorImageResolution 300 /ColorImageDepth -1 /ColorImageMinDownsampleDepth 1 /ColorImageDownsampleThreshold 1.50000 /EncodeColorImages true /ColorImageFilter /DCTEncode /AutoFilterColorImages true /ColorImageAutoFilterStrategy /JPEG /ColorACSImageDict << /QFactor 0.15 /HSamples [1 1 1 1] /VSamples [1 1 1 1] >> /ColorImageDict << /QFactor 0.15 /HSamples [1 1 1 1] /VSamples [1 1 1 1] >> /JPEG2000ColorACSImageDict << /TileWidth 256 /TileHeight 256 /Quality 30 >> /JPEG2000ColorImageDict << /TileWidth 256 /TileHeight 256 /Quality 30 >> /AntiAliasGrayImages false /CropGrayImages true /GrayImageMinResolution 300 /GrayImageMinResolutionPolicy /OK /DownsampleGrayImages true /GrayImageDownsampleType /Bicubic /GrayImageResolution 300 /GrayImageDepth -1 /GrayImageMinDownsampleDepth 2 /GrayImageDownsampleThreshold 1.50000 /EncodeGrayImages true /GrayImageFilter /DCTEncode /AutoFilterGrayImages true /GrayImageAutoFilterStrategy /JPEG /GrayACSImageDict << /QFactor 0.15 /HSamples [1 1 1 1] /VSamples [1 1 1 1] >> /GrayImageDict << /QFactor 0.15 /HSamples [1 1 1 1] /VSamples [1 1 1 1] >> /JPEG2000GrayACSImageDict << /TileWidth 256 /TileHeight 256 /Quality 30 >> /JPEG2000GrayImageDict << /TileWidth 256 /TileHeight 256 /Quality 30 >> /AntiAliasMonoImages false /CropMonoImages true /MonoImageMinResolution 1200 /MonoImageMinResolutionPolicy /OK /DownsampleMonoImages true /MonoImageDownsampleType /Bicubic /MonoImageResolution 1200 /MonoImageDepth -1 /MonoImageDownsampleThreshold 1.50000 /EncodeMonoImages true /MonoImageFilter /CCITTFaxEncode /MonoImageDict << /K -1 >> /AllowPSXObjects false /CheckCompliance [ /None ] /PDFX1aCheck false /PDFX3Check false /PDFXCompliantPDFOnly false /PDFXNoTrimBoxError true /PDFXTrimBoxToMediaBoxOffset [ 0.00000 0.00000 0.00000 0.00000 ] /PDFXSetBleedBoxToMediaBox true /PDFXBleedBoxToTrimBoxOffset [ 0.00000 0.00000 0.00000 0.00000 ] /PDFXOutputIntentProfile (None) /PDFXOutputConditionIdentifier () /PDFXOutputCondition () /PDFXRegistryName () /PDFXTrapped /False /CreateJDFFile false /Description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> /CHS <FEFF4f7f75288fd94e9b8bbe5b9a521b5efa7684002000500044004600206587686353ef901a8fc7684c976262535370673a548c002000700072006f006f00660065007200208fdb884c9ad88d2891cf62535370300260a853ef4ee54f7f75280020004100630072006f0062006100740020548c002000410064006f00620065002000520065006100640065007200200035002e003000204ee553ca66f49ad87248672c676562535f00521b5efa768400200050004400460020658768633002> /CHT <FEFF4f7f752890194e9b8a2d7f6e5efa7acb7684002000410064006f006200650020005000440046002065874ef653ef5728684c9762537088686a5f548c002000700072006f006f00660065007200204e0a73725f979ad854c18cea7684521753706548679c300260a853ef4ee54f7f75280020004100630072006f0062006100740020548c002000410064006f00620065002000520065006100640065007200200035002e003000204ee553ca66f49ad87248672c4f86958b555f5df25efa7acb76840020005000440046002065874ef63002> /CZE <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> /DAN <FEFF004200720075006700200069006e0064007300740069006c006c0069006e006700650072006e0065002000740069006c0020006100740020006f007000720065007400740065002000410064006f006200650020005000440046002d0064006f006b0075006d0065006e007400650072002000740069006c0020006b00760061006c00690074006500740073007500640073006b007200690076006e0069006e006700200065006c006c006500720020006b006f007200720065006b007400750072006c00e60073006e0069006e0067002e0020004400650020006f007000720065007400740065006400650020005000440046002d0064006f006b0075006d0065006e0074006500720020006b0061006e002000e50062006e00650073002000690020004100630072006f00620061007400200065006c006c006500720020004100630072006f006200610074002000520065006100640065007200200035002e00300020006f00670020006e0079006500720065002e> /DEU <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> /ESP <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> /ETI <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> /FRA <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> /GRE <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> /HUN <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> /ITA <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> /JPN <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> /KOR <FEFFc7740020c124c815c7440020c0acc6a9d558c5ec0020b370c2a4d06cd0d10020d504b9b0d1300020bc0f0020ad50c815ae30c5d0c11c0020ace0d488c9c8b85c0020c778c1c4d560002000410064006f0062006500200050004400460020bb38c11cb97c0020c791c131d569b2c8b2e4002e0020c774b807ac8c0020c791c131b41c00200050004400460020bb38c11cb2940020004100630072006f0062006100740020bc0f002000410064006f00620065002000520065006100640065007200200035002e00300020c774c0c1c5d0c11c0020c5f40020c2180020c788c2b5b2c8b2e4002e> /LTH <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> /LVI <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> /NLD (Gebruik deze instellingen om Adobe PDF-documenten te maken voor kwaliteitsafdrukken op desktopprinters en proofers. De gemaakte PDF-documenten kunnen worden geopend met Acrobat en Adobe Reader 5.0 en hoger.) /NOR <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> /POL <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> /PTB <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> /RUM <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> /RUS <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> /SKY <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> /SLV <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> /SUO <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> /SVE <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> /TUR <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> /UKR <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> /ENU (Use these settings to create Adobe PDF documents for quality printing on desktop printers and proofers. Created PDF documents can be opened with Acrobat and Adobe Reader 5.0 and later.) >> /Namespace [ (Adobe) (Common) (1.0) ] /OtherNamespaces [ << /AsReaderSpreads false /CropImagesToFrames true /ErrorControl /WarnAndContinue /FlattenerIgnoreSpreadOverrides false /IncludeGuidesGrids false /IncludeNonPrinting false /IncludeSlug false /Namespace [ (Adobe) (InDesign) (4.0) ] /OmitPlacedBitmaps false /OmitPlacedEPS false /OmitPlacedPDF false /SimulateOverprint /Legacy >> << /AddBleedMarks false /AddColorBars false /AddCropMarks false /AddPageInfo false /AddRegMarks false /ConvertColors /NoConversion /DestinationProfileName () /DestinationProfileSelector /NA /Downsample16BitImages true /FlattenerPreset << /PresetSelector /MediumResolution >> /FormElements false /GenerateStructure true /IncludeBookmarks false /IncludeHyperlinks false /IncludeInteractive false /IncludeLayers false /IncludeProfiles true /MultimediaHandling /UseObjectSettings /Namespace [ (Adobe) (CreativeSuite) (2.0) ] /PDFXOutputIntentProfileSelector /NA /PreserveEditing true /UntaggedCMYKHandling /LeaveUntagged /UntaggedRGBHandling /LeaveUntagged /UseDocumentBleed false >> ] >> setdistillerparams << /HWResolution [2400 2400] /PageSize [595.440 793.440] >> setpagedevice

Commentary

Ultra-processed foods: what they are and how to identify them

Carlos A Monteiro1,2,*, Geoffrey Cannon2, Renata B Levy2,3, Jean-Claude Moubarac4, Maria LC Louzada2, Fernanda Rauber2, Neha Khandpur2, Gustavo Cediel2, Daniela Neri2, Euridice Martinez-Steele2, Larissa G Baraldi2 and Patricia C Jaime1,2 1Department of Nutrition, School of Public Health, University of São Paulo, São Paulo, Brazil: 2Center for Epidemiological Research in Nutrition and Health, Department of Nutrition, School of Public Health, University of São Paulo, Av. Dr Arnaldo 715, São Paulo, SP 01246-904, Brazil: 3Department of Preventive Medicine, School of Medicine, University of São Paulo, São Paulo, Brazil: 4Département de Nutrition, Université de Montréal, Montréal, Canada

Submitted 3 September 2018: Final revision received 21 November 2018: Accepted 30 November 2018: First published online 12 February 2019

Abstract The present commentary contains a clear and simple guide designed to identify ultra-processed foods. It responds to the growing interest in ultra-processed foods among policy makers, academic researchers, health professionals, journalists and consumers concerned to devise policies, investigate dietary patterns, advise people, prepare media coverage, and when buying food and checking labels in shops or at home. Ultra-processed foods are defined within the NOVA classification system, which groups foods according to the extent and purpose of industrial processing. Processes enabling the manufacture of ultra-processed foods include the fractioning of whole foods into substances, chemical modifications of these substances, assembly of unmodified and modified food substances, frequent use of cosmetic additives and sophisticated packaging. Processes and ingredients used to manufacture ultra-processed foods are designed to create highly profitable (low-cost ingredients, long shelf-life, emphatic branding), convenient (ready-to-consume), hyper-palatable products liable to displace all other NOVA food groups, notably unprocessed or minimally processed foods. A practical way to identify an ultra-processed product is to check to see if its list of ingredients contains at least one item characteristic of the NOVA ultra-processed food group, which is to say, either food substances never or rarely used in kitchens (such as high-fructose corn syrup, hydrogenated or interesterified oils, and hydrolysed proteins), or classes of additives designed to make the final product palatable or more appealing (such as flavours, flavour enhancers, colours, emulsifiers, emulsifying salts, sweeteners, thickeners, and anti- foaming, bulking, carbonating, foaming, gelling and glazing agents).

Keywords Ultra-processed food

Food processing Food classification

NOVA

Ultra-processed foods already make up more than half of the total dietary energy consumed in high-income coun- tries such as the USA(1), Canada(2) and the UK(3) and between one-fifth and one-third of total dietary energy in middle-income countries such as Brazil(4), Mexico(5) and Chile(6). The average growth in sales of these products amounts to about 1% per year in high-income countries and up to 10% per year in middle-income countries(7).

Population-based studies conducted in several coun- tries, most of them using national dietary intake surveys,

have shown that ultra-processed foods are typically high- energy-dense products, high in sugar, unhealthy fats and salt, and low in dietary fibre, protein, vitamins and minerals(2–4,6,8–13). Experimental studies indicate that ultra- processed foods induce high glycaemic responses and have low satiety potential(14), and create a gut environ- ment that selects microbes that promote diverse forms of inflammatory disease(15). Cross-sectional and longitudinal studies have shown that increases in the dietary share of ultra-processed foods result in deterioration of the

Public Health Nutrition: 22(5), 936–941 doi:10.1017/S1368980018003762

*Corresponding author: Email [email protected] © The Authors 2019

nutritional quality of the overall diet(2–4,6–13,16–19) and increased obesity(20–23), hypertension(24), coronary and cerebrovascular diseases(25), dyslipidaemia(26), metabolic syndrome(27), gastrointestinal disorders(28), and total and breast cancer(29). Avoidance of ultra-processed foods is the ‘golden rule’ of national dietary guidelines issued recently in Latin American countries(30,31).

While much has been published on ultra-processed foods in peer-reviewed journals(1–24,26–30), reports from UN agencies(32–35) and in broadcast and written media(36–39), a simple method to identify these products has not yet been made explicit. The present commentary addresses this gap by defining ultra-processed foods within the context of the NOVA food classification sys- tem(40) and showing how they can be confidently identified.

Defining ultra-processed foods

Almost all foods are processed to some extent, if only by preservation, and it is therefore unhelpful to criticise foods as being ‘processed’. A number of food classifications have been devised that pay special attention to types of pro- cessing. A systematic review has shown that, of these, NOVA is the most specific, coherent, clear, comprehensive and workable(41).

NOVA classifies all foods and food products into four groups according to the extent and purpose of the industrial processing they undergo. It considers all physi- cal, biological and chemical methods used during the food manufacturing process, including the use of additives(40).

A summary of the types and purposes of the industrial processes that define each of the four NOVA groups, shown below, makes it easy to understand the unique features of ultra-processed foods and to appreciate the health concerns associated with their consumption. Full definitions and lists of examples of each of the four NOVA groups are provided in the online supplementary material, Supplemental Table 1.

Non-ultra-processed food groups Minimally processed foods, that together with unpro- cessed foods make up NOVA group 1, are unprocessed foods altered by industrial processes such as removal of inedible or unwanted parts, drying, crushing, grinding, fractioning, roasting, boiling, pasteurization, refrigeration, freezing, placing in containers, vacuum packaging or non- alcoholic fermentation. None of these processes add salt, sugar, oils or fats, or other food substances to the original food. Their main aim is to extend the life of grains (cer- eals), legumes (pulses), vegetables, fruits, nuts, milk, meat and other foods, enabling their storage for longer use, and often to make their preparation easier or more diverse.

NOVA group 2 is of processed culinary ingredients. These are substances obtained directly from group 1 foods or from nature, like oils and fats, sugar and salt. They are created by industrial processes such as pressing, cen- trifuging, refining, extracting or mining, and their use is in the preparation, seasoning and cooking of group 1 foods.

NOVA group 3 is of processed foods. These are industrial products made by adding salt, sugar or other substance found in group 2 to group 1 foods, using pre- servation methods such as canning and bottling, and, in the case of breads and cheeses, using non-alcoholic fer- mentation. Food processing here aims to increase the durability of group 1 foods and make them more enjoy- able by modifying or enhancing their sensory qualities.

Traditional and long-established dietary patterns all over the world, including those known to promote long and healthy lives such as those in Mediterranean countries(42), Japan(43) and Korea(44), have been and are based on dishes and meals made from a variety of unprocessed or minimally processed plant foods, prepared, seasoned and cooked with processed culinary ingredients and com- plemented with processed foods.

The ultra-processed food group Ultra-processed foods are formulations of ingredients, mostly of exclusive industrial use, that result from a series of industrial processes (hence ‘ultra-processed’).

Processes enabling the manufacture of ultra-processed foods involve several steps and different industries. It starts with the fractioning of whole foods into substances that include sugars, oils and fats, proteins, starches and fibre. These substances are often obtained from a few high-yield plant foods (corn, wheat, soya, cane or beet) and from puréeing or grinding animal carcasses, usually from intensive livestock farming. Some of these substances are then sub- mitted to hydrolysis, or hydrogenation, or other chemical modifications. Subsequent processes involve the assembly of unmodified and modified food substances with little if any whole food using industrial techniques such as extrusion, moulding and pre-frying. Colours, flavours, emulsifiers and other additives are frequently added to make the final pro- duct palatable or hyper-palatable. Processes end with sophisticated packaging usually with synthetic materials.

Sugar, oils and fats, and salt, used to make processed foods, are often ingredients of ultra-processed foods, gen- erally in combination. Additives that prolong product dura- tion, protect original properties and prevent proliferation of micro-organisms may be used in both processed and ultra- processed foods, as well as in processed culinary ingre- dients, and, infrequently, in minimally processed foods.

Ingredients that are characteristic of ultra-processed foods can be divided into food substances of no or rare culinary use and classes of additives whose function is to make the final product palatable or often hyper-palatable (‘cosmetic additives’). Food substances of no or rare

How to identify ultra-processed foods 937

culinary use, and used only in the manufacture of ultra- processed foods, include varieties of sugars (fructose, high-fructose corn syrup, ‘fruit juice concentrates’, invert sugar, maltodextrin, dextrose, lactose), modified oils (hydrogenated or interesterified oils) and protein sources (hydrolysed proteins, soya protein isolate, gluten, casein, whey protein and ‘mechanically separated meat’). Cos- metic additives, also used only in the manufacture of ultra- processed foods, are flavours, flavour enhancers, colours, emulsifiers, emulsifying salts, sweeteners, thickeners, and anti-foaming, bulking, carbonating, foaming, gelling and glazing agents. These classes of additives disguise undesirable sensory properties created by ingredients, processes or packaging used in the manufacture of ultra- processed foods, or else give the final product sensory properties especially attractive to see, taste, smell and/ or touch.

Ultra-processed foods include carbonated soft drinks; sweet or savoury packaged snacks; chocolate, candies (confectionery); ice cream; mass-produced packaged breads and buns; margarines and other spreads; cookies (biscuits), pastries, cakes and cake mixes; breakfast ‘cer- eals’; pre-prepared pies and pasta and pizza dishes; poultry and fish ‘nuggets’ and ‘sticks’, sausages, burgers, hot dogs and other reconstituted meat products; pow- dered and packaged ‘instant’ soups, noodles and desserts; and many other products (see online supplementary material, Supplemental Table 1).

Processes and ingredients used for the manufacture of ultra-processed foods are designed to create highly profitable products (low-cost ingredients, long shelf-life, branded products) which are liable to displace all other NOVA food groups. Their convenience (imperishable, ready-to-consume), hyper-palatability, branding and ownership by transnational corporations, and aggressive marketing give ultra-processed foods enormous market advantages over all other NOVA food groups. Marketing strategies used worldwide include vivid packaging, health claims, special deals with retailers to secure prime shelf space, establishment of franchised catering outlets, and campaigns using social, electronic, broadcast and print media, including to children and in schools, often with vast budgets. All this explains why ultra-processed foods have been successful in displacing unprocessed or minimally processed foods and freshly prepared dishes and meals – or ‘real food’ – in most parts of the world(7,45–47).

The nature of the processes and ingredients used in their manufacture, and their displacement of unprocessed or minimally processed foods and freshly prepared dishes and meals, make ultra-processed foods intrinsically unhealthy. The disorders and diseases associated with diets largely made up from ultra-processed foods, and the mechanisms linking these diets to specific diseases, are described elsewhere(40). The displacement of ‘real food’ by ultra-processed foods is also a cause of social, cultural,

economic, political and environmental disruption and crises. These are also described elsewhere(40).

Identifying ultra-processed foods

The food manufacturing industry is not obliged to state on food labels the processes used in its products and even less the purposes of these processes. In some cases, this can make confident identification of ultra-processed foods difficult for consumers, health professionals, policy makers and even for researchers.

There is of course no need to examine every food to know whether or not it belongs to the ultra-processed food group. As stated above, and to take a few examples, fresh vegetables, fruits, and starchy roots and tubers are obviously not ultra-processed; nor are pasteurized milk and chilled meat. Plant oils, sugar and salt, typically used in culinary preparations in combination with unprocessed or minimally processed foods, are also not ultra- processed.

It is however not always immediately clear when some specific food products are ultra-processed or not. Exam- ples include breads and breakfast cereals. Here the solu- tion is to examine the ingredients labels that by law must be included on pre-packaged food and drink products.

Industrial breads made only from wheat flour, water, salt and yeast are processed foods, while those whose lists of ingredients also include emulsifiers or colours are ultra- processed. Plain steel-cut oats, plain corn flakes and shredded wheat are minimally processed foods, while the same foods are processed when they also contain sugar, and ultra-processed if they also contain flavours or colours.

Generally, the practical way to identify if a product is ultra-processed is to check to see if its list of ingredients contains at least one item characteristic of the ultra- processed food group, which is to say, either food sub- stances never or rarely used in kitchens, or classes of additives whose function is to make the final product palatable or more appealing (‘cosmetic additives’).

Food substances not used in kitchens appear in the beginning or in the middle of the lists of ingredients of ultra-processed foods. These include hydrolysed proteins, soya protein isolate, gluten, casein, whey protein, ‘mechanically separated meat’, fructose, high-fructose corn syrup, ‘fruit juice concentrate’, invert sugar, maltodextrin, dextrose, lactose, soluble or insoluble fibre, hydrogenated or interesterified oil; and also other sources of protein, carbohydrate or fat which are neither foods from NOVA group 1 or group 3, nor culinary ingredients from NOVA group 2. The presence in the list of ingredients of one or more of these food substances identifies a product as ultra-processed.

Cosmetic additives are at the end of lists of ingredients of ultra-processed foods, together with other additives. As

938 CA Monteiro et al.

said above, cosmetic additives include flavours, flavour enhancers, colours, emulsifiers, emulsifying salts, sweet- eners, thickeners, and anti-foaming, bulking, carbonating, foaming, gelling and glazing agents. The presence in the list of ingredients of one or more additives that belong to these classes of additives also identifies a product as ultra- processed.

Although information in ingredients labels is not fully standardized in all countries, some of the most frequently used cosmetic additives such as flavours, flavour enhan- cers, colours and emulsifiers are usually easy to identify in ingredients lists. They are often expressed as a class, such as flavourings or natural flavours or artificial flavours; or their names are followed by their class, such as ‘mono- sodium glutamate (flavour enhancer)’, or ‘caramel colour’, or ‘soya lecithin as emulsifier’. Other cosmetic additives may be known to consumers, such as certain types of sweeteners like aspartame, cyclamate or compounds derived from stevia. In any case, the UN Codex Alimentarius provides a regularly updated list of additives with their functional classes(48) as well as an online search facility where both names and classes of additives can be browsed(49).

Conclusion

Most foods as purchased and consumed are processed to some extent. For this reason, accounts that are critical of ‘processed food’ are not useful. Diets restricted to unpro- cessed food would be less diverse and less secure. Foods benefit, and are made more available, when processed by various harmless methods of preservation; and some processes enhance food quality, non-alcoholic fermenta- tion being an example. Traditional and established cui- sines all over the world are based on dishes and meals prepared from unprocessed and minimally processed food together with processed culinary ingredients and pro- cessed foods. The issue is not processing. It is ultra- processed foods, the fourth group in the NOVA system of food classification.

Ultra-processed foods are not ‘real food’. As stated, they are formulations of food substances often modified by chemical processes and then assembled into ready-to- consume hyper-palatable food and drink products using flavours, colours, emulsifiers and a myriad of other cos- metic additives. Most are made and promoted by trans- national and other giant corporations. Their ultra- processing makes them highly profitable, intensely appealing and intrinsically unhealthy.

The present commentary shows how to identify ultra- processed foods, and is designed for policy makers, researchers, health professionals, journalists and con- sumers. Computer software and cell phone apps that scan and interpret food package barcodes should make this identification even easier. A cell phone app created by the

non-profit organization Open Food Facts, based in France, already enables consumers to identify among more than 145 000 packaged products the more than 75 000 that are ultra-processed(50).

Acknowledgements

Financial support: This work was supported by the Fun- dação de Amparo à Pesquisa do Estado de São Paulo (FAPESP; grant number 2015/14900-9). The conclusions and opinions in this article are those of the authors and do not necessarily reflect those of FAPESP. Conflict of interest: The authors declare that there are no conflicts of interests. Authorship: All authors participated in the con- ception and drafting of the commentary. Ethics of human subject participation: Not applicable.

Supplementary material

To view supplementary material for this article, please visit https://doi.org/10.1017/S1368980018003762

References

1. Baraldi LG, Martinez Steele EM, Canella DS et al. (2018) Consumption of ultra-processed foods and associated sociodemographic factors in the USA between 2007 and 2012: evidence from a nationally representative cross- sectional study. BMJ Open 8, e020574.

2. Moubarac J-C, Batal M, Louzada ML et al. (2017) Con- sumption of ultra-processed foods predicts diet quality in Canada. Appetite 108, 512–520.

3. Rauber F, Louzada MLC, Steele EM et al. (2018) Ultra- processed food consumption and chronic non- communicable diseases-related dietary nutrient profile in the UK (2008–2014). Nutrients 10, E587.

4. Louzada MLDC, Ricardo CZ, Steele EM et al. (2018) The share of ultra-processed foods determines the overall nutritional quality of diets in Brazil. Public Health Nutr 21, 94–102.

5. Marrón-Ponce JA, Sánchez-Pimienta TG, Louzada M et al. (2018) Energy contribution of NOVA food groups and sociodemographic determinants of ultra-processed food consumption in the Mexican population. Public Health Nutr 21, 87–93.

6. Cediel G, Reyes M, da Costa Louzada ML et al. (2018) Ultra- processed foods and added sugars in the Chilean diet (2010). Public Health Nutr 21, 125–133.

7. Monteiro CA, Moubarac JC, Cannon G et al. (2013) Ultra- processed products are becoming dominant in the global food system. Obes Rev 14, 21–28.

8. Louzada MLC, Martins APB, Canella DS et al. (2015) Impact of ultra-processed foods on micronutrient content in the Brazilian diet. Rev Saude Publica 49, 45.

9. Costa Louzada ML, Martins AP, Canella DS et al. (2015) Ultra-processed foods and the nutritional dietary profile in Brazil. Rev Saude Publica 45, 38.

10. Monteiro CA, Levy RB, Claro RM et al. (2011) Increasing consumption of ultra-processed foods and likely impact on human health: evidence from Brazil. Public Health Nutr 14, 5–13.

How to identify ultra-processed foods 939

11. Moubarac JC, Martins AP, Claro RM et al. (2013) Con- sumption of ultra-processed foods and likely impact on human health. Evidence from Canada. Public Health Nutr 16, 2240–2248.

12. Poti JM, Mendez MA, Ng SW et al. (2015) Is the degree of food processing and convenience linked with the nutritional quality of foods purchased by US households? Am J Clin Nutr 101, 1251–1262.

13. Luiten CM, Steenhuis IH, Eyles H et al. (2016) Ultra- processed foods have the worst nutrient profile, yet they are the most available packaged products in a sample of New Zealand supermarkets. Public Health Nutr 19, 530–538.

14. Fardet A (2016) Minimally processed foods are more satiating and less hyperglycemic than ultra-processed foods: a preliminary study with 98 ready-to-eat foods. Food Funct 7, 2338–2346.

15. Zinöcker MK & Lindseth IA (2018) The Western diet– microbiome–host interaction and its role in metabolic dis- ease. Nutrients 10, E365.

16. Martinez Steele E, Baraldi LG, Louzada ML et al. (2016) Ultra-processed foods and added sugars in the US diet: evidence from a nationally representative cross- sectional study. BMJ Open 6, e009892.

17. Julia C, Martinez L, Allès B et al. (2018) Contribution of ultra- processed foods in the diet of adults from the French NutriNet-Santé study. Public Health Nutr 21, 27–37.

18. Martínez Steele E, Raubenheimer D, Simpson SJ et al. (2018) Ultra-processed foods, protein leverage and energy intake in the USA. Public Health Nutr 21, 114–124.

19. Steele EM, Popkin B, Swinburn B et al. (2017) The share of ultra-processed foods and the overall nutritional quality of diets in the US: evidence from a nationally representative cross-sectional study. Popul Health Metr 15, 6.

20. Louzada MLC, Baraldi LC, Steele EM et al. (2015) Con- sumption of ultra-processed foods and obesity in Brazilian adolescents and adults. Prev Med 81, 8–15.

21. Mendonça R, Pimenta A, Gea A et al. (2016) Ultra-processed foods consumption and risk of overweight/obesity: the SUN cohort study. Am J Clin Nutr 104, 1433–1440.

22. Filippa Juul F, Steele EM, Parekh N et al. (2018) Ultra- processed food consumption and excess weight among US adults. Br J Nutr 120, 90–100.

23. Nardocci M, Leclerc B-S, Louzada ML et al. (2019) Con- sumption of ultra-processed foods and obesity in Canada. Can J Public Health 110, 4–14.

24. Mendonça R, Lopes A, Pimenta AM et al. (2017) Ultra- processed food consumption and the incidence of hyper- tension in a Mediterranean cohort: the Seguimiento Universidad de Navarra Project. Am J Hypertens 30, 358–366.

25. Srour B, Fezeu L, Kesse-Guyot E et al. (2018) Consomma- tion d’aliments ultra-transformés et risque de maladies car- diovasculaires dans la cohorte NutriNet-Santé. In Livre des abstracts, Journée Francophone de Nutrition, Nice, France, 28–30 Novembre 2018, p. 61. Levallois-Perret: JFN 2018/ MCI France.

26. Rauber F, Campagnolo PDB, Hoffman DJ et al. (2015) Consumption of ultra-processed food products and its effects on children’s lipid profiles: a longitudinal study. Nutr Metab Cardiovasc Dis 25, 116–122.

27. Lavigne-Robichaud M, Moubarac J-C, Lantagne-Lopez S et al. (2018) Diet quality indices in relation to metabolic syndrome in an Indigenous Cree (Eeyouch) population in northern Québec, Canada. Public Health Nutr 21, 172–180.

28. Schnabel L, Buscail C, Sabate J-M et al. (2018) Association between ultra-processed food consumption and functional gastrointestinal disorders: results from the French NutriNet- Santé Cohort. Am J Gastroenterol 113, 1217–1228.

29. Fiolet T, Srour B, Sellem L et al. (2018) Consumption of ultra-processed foods and cancer risk: results from NutriNet- Santé prospective cohort. BMJ 360, k322.

30. Monteiro CA, Cannon G, Moubarac J-C et al. (2015) Dietary guidelines to nourish humanity and the planet in the twenty-first century. A blueprint from Brazil. Public Health Nutr 18, 2311–2322.

31. Ministerio de Salud (2016) Guía Alimentaria para la Población Uruguaya. http://msp.gub.uy/sites/default/files/ archivos_adjuntos/MS_guia_web.pdf (accessed November 2018).

32. Pan American Health Organization (2015) Ultra-Processed Food and Drink Products in Latin America: Trends, Impact on Obesity, Policy Implications. Washington, DC: PAHO; available at http://www.paho.org/hq/index.php?option= com_content&view= article&id= 11153%3Aultra-processed- food-and-drinkproducts&catid=4999%3Adocuments&lang=en

33. Food and Agriculture Organization of the United Nations (2015) Guidelines on the Collection of Information on Food Processing Through Food Consumption Surveys. Rome: FAO; available at http://www.fao.org/3/a-i4690e.pdf

34. Food and Agriculture Organization of the United Nations & Food Climate Research Network (2016) Plates, Pyramids, Planet. Developments in National Healthy and Sustainable Dietary Guidelines: A State of Play Assessment. Rome/ Oxford: FAO/University of Oxford; available at http://www.fao.org/3/a-i5640e.pdf

35. Pan American Health Organization (2018) Ultra-Processed Food and Drink Products in Latin America: Sales, Sources, Nutrient Profiles and Policy Implications. Washington, DC: PAHO (In the Press).

36. Nestle M (2010) How ultra-processed foods are killing us. The Atlantic, 4 November. https://www.theatlantic.com/ health/archive/2010/11/how-ultra-processed-foods-are-killing- us/65614/ (accessed November 2018).

37. Welch A (2016) ‘Ultra-processed foods’ a huge chunk of American diet. CBS News, 10 March. https://www.cbsnews. com/news/ultra-processed-foods-calories-american-diet/ (accessed November 2018).

38. Boseley S (2018) ‘Ultra-processed’ products now half of all UK family food purchases. The Guardian, 2 February. https://www.theguardian.com/science/2018/feb/02/ultra- processed-products-now-half-of-all-uk-family-food-purchases (accessed November 2018).

39. Boseley S (2018) Ultra-processed foods may be linked to cancer, says study. The Guardian, 14 February. https:// www.theguardian.com/science/2018/feb/14/ultra-processed- foods-may-be-linked-to-cancer-says-study (accessed Novem- ber 2018).

40. Monteiro CA, Cannon G, Moubarac J-C et al. (2018) The UN Decade of Nutrition, the NOVA food classification and the trouble with ultra-processing. Public Health Nutr 21, 5–17.

41. Moubarac J-C, Parra D, Cannon G et al. (2014) Food classification systems based on food processing: sig- nificance and implications for policies and actions: a systematic literature review and assessment. Curr Obes Rep 3, 256–272.

42. Sofi F, Abbate R, Gensini GF et al. (2010) Accruing evidence on benefits of adherence to the Mediterranean diet on health: an updated systematic review and meta-analysis. Am J Clin Nutr 92, 1189–1196.

43. Sho H (2001) History and characteristics of Okinawan longevity food. Asia Pac J Clin Nutr 10, 159–164.

44. Jung SJ, Park SH, Choi EK et al. (2014) Beneficial effects of Korean traditional diets in hypertensive and type 2 diabetic patients. J Med Food 17, 161–171.

45. Moubarac JC, Batal M, Martins AP et al. (2014) Processed and ultra-processed food products: consumption trends in Canada from 1938 to 2011. Can J Diet Pract Res 75, 15–21.

940 CA Monteiro et al.

46. Monteiro CA & Cannon G (2012) The impact of transna- tional ‘big food’ companies on the South: a view from Brazil. PLoS Med 9, e1001252.

47. Moodie R, Stuckler D, Monteiro CA et al. (2013) Profits and pandemics: prevention of harmful effects of tobacco, alcohol, and ultra-processed food and drink industries. Non-communicable Diseases Series 4. Lancet 381, 670–679.

48. Food and Agriculture Organization & World Health Organization (2017) International Food Standards. Codex Alimentarius. Class and Names and the International

Numbering System for Food Additives. Adopted 1989. Revision 2008. Amendment: FAO/WHO. http://www.fao. org/tempref/codex/Meetings/CCFAC/ccfac31/INS_e.pdf (accessed November 2018).

49. Food and Agriculture Organization & World Health Orga- nization (n.d.) FAO/WHO Food Standards. Codex Ali- mentarius. GSFA online. http://www.fao.org/gsfaonline/ additives/results.html (accessed November 2018).

50. Open Food Facts (2018) List of NOVA groups – World. https://world.openfoodfacts.org/nova-groups (accessed November 2018).

How to identify ultra-processed foods 941