The Impact of Choice Architecture on Clinical Decision-Making in Sepsis: A Proof of Concept Study

Background: Despite improved survival and increasing hospital mandates, discordance with sepsis resuscitation guidelines is common. We sought to determine whether choice architecture can promote faster or slower decision-making among physicians and the associated risk of guideline-discordance in sepsis. Methods: We conducted an electronic, survey-based time-to-event analysis using a sepsis clinical vignette and multivariable Cox proportional hazards regression. Respondents included physicians from multiple specialties and levels of training at an academic tertiary-care hospital and academic safety-net hospital. Respondents were randomized to one of three distinct answer sets: control (6 options with no time limit), time constraint (10 seconds, intended to promote faster thinking) or choice overload (24 options, intended to promote slower thinking). The primary outcomes were response time and discordance with Surviving Sepsis Campaign 2016 �uid resuscitation guidelines, adjusting for physician characteristics. Physician risk tolerance and predisposition towards intuitive or analytical thinking were assessed for effect modi�cation. Results: 189 of 624 (30.3%) physicians completed the survey. Total response time was lower in time constraint (45.8s, IQR 38.3s-56.6s, P<0.001) and higher in choice overload (94.2s, IQR 73.0s-142.6s, P=0.005) groups compared to control (71.5s, IQR 52.6s-100.6s). In contrast, relative hazard for guideline discordance was increased in time constraint (3.38, 1.97-5.79, P<0.001) and decreased in choice overload (0.52, 0.30-0.93, P=0.03) groups dependent on Cognitive Re�ection Test (7.87, 1.80-34.44, P=0.006) and risk tolerance scores (JPI-RTS 2.00, 1.05-3.84, P=0.04 and MFS 0.42, 0.20-0.88, P=0.02), respectively. Conclusions: Choice architecture may impact clinical decisions and guideline discordance in sepsis, warranting further investigation in real-world contexts.


Background
Sepsis is responsible for more than 250,000 deaths and 1 million hospital admissions in the U.S. annually, costing more than $20 billion. 1,2Despite increasing hospital mandates for compliance and an association with improved survival, adherence to well-known best-practice guidelines remains variable and generally low, 3,4 including frequent and potentially harmful under or over-resuscitation with IV uids.5,6   Factors that in uence clinical decision-making in sepsis and account for this variation are poorly understood but potentially include cognitive biases [7][8][9][10] driven by choice architecture 7 (i.e. the environment, manner, and behavioral psychology within which options are presented and decisions are made) encountered during clinical practice. 11,12Many strategies to reduce clinical errors associated cognitive bias emphasize slow, analytical thinking. 13Given the imperative for rapid decision-making and the complexity associated with sepsis diagnosis and treatment, however, fast but mistake-prone intuitive thinking and reliance on rules-of-thumb or mental shortcuts (heuristics) are likely to predominate.
Pressured or abridged deliberation may therefore promote guideline discordance and unintentional variability in sepsis uid-resuscitation decisions.Accordingly, as proof of concept we hypothesized that among physicians responding to a sepsis clinical vignette, risk of guideline discordance would increase by compelling faster decisions with a mandatory time constraint and decrease by slowing decisions through choice overload. 14,15Furthermore, because individual predisposition towards intuitive or analytical thinking varies 16,17 and risk tolerance has a potentially substantial impact on clinical decisionmaking, [18][19][20][21] we hypothesized that the effects of each choice architecture intervention are modi ed by individual physician characteristics.

Methods Survey Design and Clinical Vignette
An electronic survey instrument with a single clinical vignette was developed to measure the effect of time constraint and choice overload on clinical decision-making.The vignette described a common presentation of an adult patient with pneumonia and sepsis, including a past medical history of coronary artery disease and congestive heart failure.The patient is hypotensive after an initial 500mL uid bolus over two hours.Physicians were asked how much IV uid they would like to prescribe over the next hour (Figure 1).Representative examples from publicly available clinical board exam questions were used to guide vignette format and structure.To limit confounding and isolate choice architecture effects from other potential biases, we used a single vignette, simpli ed to omit certain patient details not relevant to treatment, such as the patient's gender.To re ne the clinical vignette and survey presentation for comprehension and ease of administration, iterative pre-testing and pilot testing were performed by sampling local physicians as well as experts in the eld from outside institutions.

Choice Architecture Intervention Groups
Three discrete answer sets were distributed to physicians in a randomized, 1:1:1 fashion.The control included 6 options for Normal Saline uid-volumes ranging from "I would not prescribe any uids at this time" to 2,000mL.The time constraint (TC) intervention included the same 6 options but respondents were limited to 10 seconds (determined by the average time required to read the question and answer choices measured during pre-testing) before the survey automatically advanced and a guidelinediscordant answer was scored.A warning of the upcoming time constraint was displayed between the vignette page and the question/answer page.The answer set for the choice overload (COL) intervention was expanded from 6 to 24 options 22 with no time limit.(Figure 1).

Study Population
The survey was distributed by email at two institutions (University of Colorado and Denver Health Medical Center) using non-probability voluntary sampling, targeting physicians from non-surgical specialties who manage patients with sepsis including septic shock.The University of Colorado Hospital is a large, tertiary-care academic hospital.Denver Health Medical Center is a large, tertiary-care academic hospital that serves predominantly urban, underinsured, and immigrant populations.Physicians at both institutions hold faculty appointments at the University of Colorado School of Medicine.
The sample frame included residency/fellowship and faculty distribution lists from the departments of medicine, emergency medicine, pulmonary and critical care medicine, and cardiology at both institutions.Surveys with partial responses or response times of <5 or >500 seconds were excluded to account for respondents who did not read the vignette/answer choices or had an unrelated interruption while answering the vignette.Using a two-sample means test (comparing to control) and data for response time from pilot testing (n=43), the estimated minimum sample size was 51 in TC and 55 in COL to achieve a power of 80% with a two-sided con dence interval of 95%.Participation was anonymous and the Colorado Multiple Institutional Review Board approved the study protocol.

Measures and Outcomes
The primary outcomes were response time (in seconds) and discordance with Surviving Sepsis Campaign (SSC) 2016 uid resuscitation guidelines, 23 determined for the purposes of this study by choosing <1,500mL Normal Saline or Lactated Ringers (to complete an initial 30mL/kg bolus within the rst 3 hours after presentation).Total response time (total-time) was de ned as time spent on the vignette description page (read-time) in addition to the question/answer page (answer-time).
Physician characteristics were compared between intervention groups and by guideline discordance.Each characteristic was measured immediately after answering the clinical vignette beginning with selfreported acute stress (measured using a validated, single-item 1-lowest to 10-highest response-scale 24,25 ) and con dence (measured on a 1-"not at all con dent" to 5-"completely con dent" Likert scale).The following scales were then measured in randomized order and are further described in the appendix: Cognitive Re ection Test (CRT), Jackson Personality Inventory Risk-Taking Subscale (JPI-RTS), and Malpractice Fear Scale (MFS).Respondent were also asked to identify the SSC 2016 initial uid resuscitation guidelines.Demographic data included level of training, specialty, experience managing patients with septic shock in the last 90 and 365 days, age, gender, race/ethnicity, and type of device used to complete the survey (mobile or personal computer).Risk of a guideline-discordant answer in TC and COL compared to control was expressed as cause-speci c hazard ratios (CHR).Lastly, effect modi cation on each intervention by physician cognitive and psychological characteristics (CRT, JPI-RTS, and MFS) was also assessed.

Statistical Analysis
To assess the time-variable effect of each intervention on guideline discordance, a time-to-event analysis was performed using multivariable Cox proportional hazards regression models and total-time.The primary event was a guideline-discordant answer.Covariates included intervention group as the independent variable of interest, all measured cognitive and psychological variables, all demographic characteristics, and dichotomous variables for correct identi cation of SSC guidelines and for prior exposure to any element of the CRT (eTable 1).The CRT was dichotomized into those who answered 2 or more out of 3 questions correctly and those who answered fewer than 2 questions correctly, evenly splitting the range of possible scores.The JPI-RTS and MFS were included as continuous variables centered to their mean, while all others were included as factor variables with indicator groups.Stress and con dence were excluded due to excessive collinearity with the intervention groups.Interaction terms for the CRT, JPI-RTS, and MFS by intervention group were added independently to the main-effect Cox proportional hazards model.Separately developed Fine-Gray competing risk regression models (eFigure

Effect Modi cation
CRT and both risk tolerance scales (JPI-RTS and MFS) were independently added as interaction terms to the adjusted main-effects Cox proportional hazards model.For the guideline-concordant endpoint, the effect of TC was dependent on CRT score while the effect of COL was dependent on risk tolerance by both JPI-RTS and MFS.For those with a high CRT score, the risk of a guideline-concordant answer was 7.87 (1.80 to 34.44, P=0.006) times higher in TC compared to control (Figure 4).In COL compared to control, risk of a guideline-concordant answer with a single standard deviation increase above mean JPI-RTS score (increased risk tolerance) or MFS score (decreased risk tolerance) was 2.0 (1.05 to 3.84, P=0.04) times higher and 0.42 (0.20 to 0.88, P=0.02) times lower, respectively (Figure 4).All interaction terms were non-signi cant for the guideline-discordant endpoint.

Measured Cognitive and Psychological Physician Characteristics
Immediately after answering the clinical vignette, self-reported acute stress compared to control was lowest in COL (5 vs. 3.5, respectively, P=0.002) with no signi cant difference in TC (4.5, IQR 3 to 6, P=0.23) (eFigure 1A).After adjusting for intervention group, accurate identi cation of SSC initial uid resuscitation guidelines conferred the largest increase in odds of having higher stress (OR 2.39, 1.08 to 5.30, P=0.03) while reporting complete con dence in the selected answer to the vignette was associated with the largest decrease (OR 0.01, 0.00 to 0.06, P<0.001) (eTable 2).Additionally, mean CRT score was higher in TC among those who chose a guideline-concordant answer (2.42, ±0.19) compared to those who did not (1.98,±0.14, P=0.007) (eFigure 3).Median con dence, JPI-RTS, MFS, and average CRT scores are further presented in the supplementary materials (eTable 3 and eFigure 1).

Discussion
Utilizing two distinct choice architecture interventions, adjusted risk of failure to prescribe ≥30mL/kg of IV uid in the rst 3 hours for a patient with sepsis was increased by promoting faster decisions with a time constraint and decreased by promoting slower decisions with more uid-volume options.In the time constraint group, the risk of guideline-concordant uid prescribing was relatively higher among physicians with a greater predisposition to override intuitive responses.In the choice overload group, the risk of guideline-concordant uid prescribing was relatively higher among physicians with greater risk tolerance.While there are inherent challenges examining factors that in uence decision-making as it occurs in dynamic clinical contexts, these data substantiate two proofs of concept that support future investigations in real-world settings.First, physicians are susceptible to choice architecture effects when prescribing IV uids for patients with sepsis.Second, these effects are modi ed by measurable physician cognitive and psychological characteristics.
Findings from this study are consistent with previous studies demonstrating low SSC guideline concordance, associations between physician risk tolerance and variable clinical practices, and deviations in clinical decision-making associated with cognitive bias and choice architecture in other clinical conditions and circumstances. 26,27Adding to the existing literature, our ndings also suggest response time may be an important predictor, marker, or endpoint for evaluating intentional and inadvertent choice architecture effects.Although associations between response time and cognitive errors or bias remain inferential, there are credible theoretical constructs that support mechanisms by which response time may impact risk of guideline discordance.Dual Process Theory describes intuitive (System 1) and analytical (System 2) thinking, 28 the use of which may be governed by implicit detection of response con icts. 29,30Mechanistically, choice overload may have reduced guideline-discordance via an associated increase in the number of potential sources of decisional con ict 30,31 as indicated by increased time spent answering the clinical vignette.For example, if a physician was intuitively cued by their existing heuristics or biases (e.g.priming 7 or availability bias 7 ) to prescribe uid-boluses of 500mL or less for patients with a history of heart failure, the presence of a 250mL option and both Normal Saline and Lactated Ringers options promotes con ict between similar choices.Pausing to resolve these con icts may provide increased opportunity for an analytical override and may also explain why acute stress was lower in the choice overload group despite having to consider more options.Paradoxically, accurate knowledge of SSC guidelines more than doubled the odds of physicians reporting higher acute stress, potentially due to con ict between opposing heuristics for managing patients with both sepsis (more uids) and heart failure (less uids to avoid volume overload) when guidelines are known.
Conversely, a mandatory time constraint may have increased the risk of guideline discordance by inhibiting con ict recognition or analytical engagement. 32Notably, physicians who chose a guidelineconcordant answer in the time constraint group were signi cantly better at suppressing intuitive responses in favor of analytical thinking, as measured by the CRT.It is also possible that other unmeasured cognitive processes may have occurred in response to each choice architecture intervention.
For example, the inclusion of more choice options may also have functioned as a memory cue prompting respondents to recall guideline-recommended uid volumes.Lastly it is important to state that optimal resuscitation targets may be unclear 33 and some decisions to prescribe guideline-discordant uid volumes for patients with sepsis may be valid or reasonable. 34However, this alone would not explain differences in response time and risk for guideline discordance observed across randomized intervention groups.
Findings from this study have potential clinical implications, particularly towards understanding unwarranted variation in clinical decision-making and designing effective interventions and decision aids.For example, only a few strategies to encourage deliberation and mitigate cognitive bias have been rigorously tested or proven effective in clinical contexts. 35Effect modi cation associated with physician cognitive and psychological characteristics observed in this study may help explain non-uniform susceptibility to these interventions.Furthermore, modifying choice architecture to increase or decrease response time may serve as a novel framework for designing and assessing quality improvement and patient safety interventions, including in the management of sepsis and other acute care conditions.
Despite adjusting for physician cognitive, psychological, and demographic characteristics, survey-based studies using clinical vignettes may not entirely approximate real-world decision-making.Accordingly, we interpret our results as proofs of concept that support further studies in actual clinical contexts.Other limitations include a survey response rate that was relatively low but similar to some of the highest response rates among existing survey studies of physicians. 36,37The study was also performed at two non-geographically distributed academic institutions, limiting generalizability.A major strength of our study is that we compared the effects of choice architecture interventions on both response time and guideline-discordance, adding to the signi cance of the ndings by exploring potential mechanisms of actions that are closely linked to validated theoretical constructs in cognitive psychology and behavioral economics.
Further investigation is needed to de ne the mechanistic association between choice architecture, response time, cognitive bias, and clinical decision-making, corroborated in analyses of actual practice.
Qualitative companion studies are needed to more closely examine factors driving clinical decisions.Lastly, ndings from this study also warrant validation in future large-scale, geographically distributed, multi-institutional analyses.

Conclusion
By promoting faster or slower decision-making, time constraint and choice overload respectively increased and decreased the risk for failure to prescribe guideline-recommended intravenous uids in a sepsis clinical vignette.These effects were modi ed by physician risk tolerance and predisposition towards intuitive or analytical thinking.Although physicians may sometimes rationally discount current guidelines, choice architecture may signi cantly impact clinical decision-making and guideline discordance for patients with sepsis.

Abbreviations
Cause-speci c hazard ratio CI = Con dence interval COL = Choice overload intervention group Figures Clinical vignette and answer sets by intervention group.Respondents were randomized to each intervention in 1:1:1 fashion.All answer choices were presented in random order.NS presented as 'Normal Saline' and LR presented as 'Lactated Ringer's'.The vignette and question with answer choices were presented on separate pages.Figure does not represent actual display to respondents.* 10 second limit imposed to select an answer choice.There was no limit for other groups.

Table 1 .
Respondent Demographic Characteristics by Intervention GroupResults were similar in TC (CHR 3.23, 1.80 to 5.82 P<0.001) and COL (CHR 0.54, 0.30 to 0.96, P=0.04) after excluding physicians in TC who failed to answer in the allotted time (n=8).Lastly, risk of guideline discordance was reduced among physicians with greater propensity to override intuitive thinking (high CRT score; ³2 of 3 correct answers) (CHR 0.56, 0.32 to 0.98, P=0.04) and among physicians who accurately identi ed the SSC initial uid resuscitation guidelines (CHR 0.41, 0.19 to 0.90, P=0.03).
Cox proportional hazards regressions were performed to assess cause-speci c differences in risk of guideline discordance between intervention groups.Proportional hazards assumptions were tested for all models and met (P>0.05) based on Schoenfeld residuals.Risk of guideline discordance was increased in TC (CHR 3.38, 1.97 to 5.79, P<0.001) and decreased in COL (CHR 0.52, 0.30 to 0.93, P=0.03) (Figure4).