Predicting neurobehavioral performance of resident physicians in a Randomized Order Safety Trial Evaluating Resident-physician Schedules (ROSTERS)

Abstract

Objectives:

Mathematical models of human neurobehavioral performance that include the effects of acute and chronic sleep restriction can be key tools in assessment and comparison of work schedules, allowing quantitative predictions of performance when empirical assessment is impractical.

Methods:

Using such a model, we tested the hypothesis that resident physicians working an extended duration work roster (EDWR), including 24–28 hours of continuous duty and up to 88 hours per week averaged over 4 weeks, would have worse predicted performance than resident physicians working a rapidly cycling work roster (RCWR) intervention designed to reduce the duration of extended shifts. The performance metric used was attentional failures (i.e., PVT lapses). Model input was 169 actual work and sleep schedules. Outcomes were predicted hours per week during work hours spent at moderate (equivalent to 16–20 hours of continuous wakefulness) or high (equivalent to ≥20 hours of continuous wakefulness) performance impairment.

Results:

The model predicted that resident physicians working EDWR would spend significantly more time at moderate impairment (p=0.02, effect size=0.2) than those working RCWR; this difference was most pronounced during the circadian night (p<0.001). On both schedules, performance was predicted to decline from weeks 1+2 to weeks 3+4 (p<0.001), but the rate of decline was significantly greater on EDWR (p<0.01). Predicted performance impairment was inversely related to prior sleep duration (p<0.001).

Conclusions:

These findings demonstrate the utility of a mathematical model to evaluate the predicted performance profile of schedules for resident physicians and others who experience chronic sleep restriction and circadian misalignment.

Introduction

When individuals perform with during extended time awake (“sleep deprivation”), after multiple nights of insufficient sleep (“chronic sleep restriction”), and/or at an adverse circadian phase, their neurobehavioral performance is impaired^1,2,3,4. In the context of occupational health, these impairments are associated with increased risk of errors, workplace injuries, and motor vehicle crashes during the commute home^5–10. Work schedules that require long continuous duty hours and/or night shift work should therefore be designed with sleep and circadian factors in mind. Mathematical models are one tool to facilitate designing these schedules.

The US medical profession requires resident physicians to work extended duration work rosters (EDWRs) that pose challenges for human sleep and circadian biology. Currently standard schedules permit resident physicians to work 80 hours per week (averaged over 4 weeks) including continuous work shifts of 24–28 hours (‘on-call’ shift). This work pattern results in sleep deprivation, chronic sleep restriction and work during adverse circadian phases (i.e., at times when the circadian system is promoting sleep)¹¹. Not surprisingly, EDWRs have been shown to be associated with impairment in medical and neurobehavioral performance^12–14. When similarly challenging schedules are simulated in laboratory environments, they induce severe neurobehavioral (both objective performance and mood) impairment and metabolic dysfunction^15–19.

To mitigate the increased risk to both patient and physician safety^6,20,21, schedules have been developed that reduce the number of continuous duty hours, weekly work hours, and/or the frequency of extended-duration work shifts^22,23. One of these schedules was a Rapidly Cycling Work Roster (RCWR) with scheduled continuous duty hours limited to 16 h (RCR-16) consisting of a short day shift (7:00–15:00), a long day shift (7:00–22:00), and a long night shift (21:00–13:00).

The RCWR was tested against a traditional EDWR in a clinical trial design detailed below. Relative to a traditional EDWR, the RCWR intervention was found to significantly increase total sleep time and decrease incidence of serious medical errors when applied to post-graduate year 1 (PGY-1) resident physicians working in the ICU^11,14.

As a complement to quantitative experimental approaches, mathematical models have been widely used for schedule assessment, design, and optimization^{24–27,28–35}. Modeling of performance for individuals and teams²⁶ is valuable for several reasons: (i) modeling allows completely unbiased comparisons of performance between schedules, eliminating any concerns around blinding or expectancy effects; (ii) modeling allows performance to be interrogated at all times, even when it would be logistically difficult to empirically assess (e.g., while participants are performing work tasks or driving); (iii) modeling allows quantitative predictions of performance across multiple possible schedules since empirical assessment of every possible schedule would be impractical; (iv) the model can provide insights into the basis for performance differences between schedules, illuminating a path to effective interventions; and (v) once a model has appraised existing schedules, it can be used in future to predictively assess potential alternative schedules in a safe, relatively inexpensive, and rapid fashion, allowing candidate schedules to be assessed prior to a randomized control trial.

To make accurate predictions on resident physician schedules, a model should ideally be able to predict changes in neurobehavioral performance due to a combination of: (i) the dynamics of the circadian pacemaker (circadian phase is potentially shifted by rotating schedules); (ii) the effects of acute sleep deprivation (since 24–28 h shifts are permitted by the Accreditation Council for Graduate Medical Education [ACGME]); and (iii) the effects of variable timing and chronic sleep restriction that causes accumulated sleep debt over longer timescales. To date, only one model has been applied to resident physician schedules²⁶; the model included a dynamic circadian pacemaker and the effects of acute sleep deprivation on performance³⁶, but did not account for effects of chronic sleep restriction. The model also was only applied to theoretical schedules. Here, we employ a validated model of the human circadian pacemaker³³ combined with a sleep-homeostatic model that has previously been trained and tested using acute sleep deprivation and both stable and variable sleep timing within chronic sleep restriction protocols^19,37. This model was used to simulate 169 actual EDWR and RCWR schedules to test the hypothesis that RCWR improves neurobehavioral performance compared to EDWR.

Methods

The model is briefly reviewed below, followed by a description of the schedules and their implementation in the model.

Mathematical model

We recently developed a model of human neurobehavioral performance with equations based on the structure of the sleep-promoting adenosine system in the brain³⁷. The model includes both circadian and sleep homeostatic variations in performance, including both acute and chronic effects of sleep loss and recovery. Specifically, the model includes a drive,

$$D=H+aC,$$ (1)

where $H$ is the homeostatic sleep drive (based on the concentration of bound adenosine A₁ receptors), $a$ is a constant, and $C$ is the circadian rhythm. The model was previously calibrated against Psychomotor Vigilance Task (PVT) data in healthy young adults under conditions of acute sleep deprivation, chronic sleep restriction, and recovery³⁸. The PVT is a visual vigilance task that is sensitive to sleep deprivation, sleep restriction, and circadian phase^16,38,39 and has been correlated with performance levels associated with alcohol intoxication^40,41. The model generates an estimate of the number of PVT lapses (i.e., responses slower than 500 ms) that would occur during a standard 10-minute test, using a sigmoid function,

$$P=S\left(D\right).$$ (2)

The model has to date been used to simulate schedules with night sleep opportunities only^19,37. Thus, it was previously valid to assume entrainment (i.e., a relatively stable circadian phase), as has been done in other sleepiness and performance models^27,42: $C$ was modeled as a sinusoid (i.e., a constant amplitude, 24-hour rhythm). Since the model is used here to simulate schedules that involve both day and night sleep opportunities, and included light/dark patterns known to affect timing (“phase”) of the circadian pacemaker, it is important to incorporate into the performance model a dynamic model of the circadian pacemaker (i.e., a model that can accurately predict changes in phase or amplitude of the circadian rhythm in response to a light stimulus) rather than a static model (e.g., sinusoidal representation of the circadian pacemaker). For this, we followed an approach used previously in which a dynamic circadian pacemaker model is substituted in place of a sinusoidal rhythm^43,44. Specifically, the sinusoid is replaced by the best-fitting output of the latest validated version of the Jewett-Kronauer-Forger model of the human circadian pacemaker³³. This was achieved by setting

$$C=-0.608x+0.685x_c+0.075,$$ (3)

where $x$ and $x_c$ are the two pacemaker outputs of the differential-equation-based dynamic circadian pacemaker model³³. The coefficients were selected based on a least-squares fit to the previously used sinusoid.

The model used two inputs: each resident’s recorded sleep/wake patterns, and a modeled light input to the dynamic circadian pacemaker model (measured in lux, method detailed below). We used the model to generate predictions of number of attentional failures (i.e., PVT lapses) at each point in time as output.

Schedules

The schedules used for modeling were created from daily logs of sleep and work (detailed below) of resident physicians from July 2013 to March 2017 as part of the Randomized Order Safety Trial Evaluating Resident Schedules (ROSTERS) study⁴⁵. This was a multi-center cluster-randomized, clinical trial designed to evaluate the effectiveness of replacing EDWRs with RCWRs in six pediatric intensive care units (PICUs). Each condition had a 4-month wash-in period followed by an 8-month data collection period, with the order of the schedule condition being randomly assigned⁴⁵. All post-graduate year 2 (PGY-2) and 3 (PGY-3) resident physicians working in the PICU at each site over the study period were invited to participate. Full details of the study, including PVT results, are reported elsewhere^45–47. The study was approved by the Institutional Review Board (IRB) at each academic medical center as well as at the University of California San Francisco (Data Coordinating Center, DCC) and by the Partners Human Research Committee (Clinical Coordinating Center, CCC), and is reported on ClinicalTrials.gov (NCT02134847).

A daily electronic log was used to collect data from participants on sleep timing (including sleep latency estimates) and duration, awakenings, and start and end times of work schedules (“rosters”). A paper version of the same sleep log was previously shown to have high epoch-to-epoch agreement with continuous polysomnography (95.6% epoch-to-epoch agreement; r=0.94, p<0.001)¹¹. Analyses of experimental sleep and work schedule data are presented in Barger et al.⁶ For inclusion in the model, we required at least 3 consecutive weeks of sleep/wake and work data with no missing intervals. A total of 298 schedules had 3 consecutive weeks of sleep/wake and work data collection, of which 169 (57%) had no missing intervals and were suitable for modeling: 95 EDWRs and 74 RCWRs. On logs where individuals failed to report sleep latency (5.0% of nights), that same individual’s median reported sleep latency was imputed using median substitution. Schedule data included in the analysis ranged in length from 21.2 days to 33.3 days for EDWR and from 21.2 days to 33.8 days for RCWR.

Simulations

Sleep/wake schedules were implemented by forcing the model to be in sleep or wakefulness during the times recorded in the logs. Light levels were simulated using a previously published method for simulating how individuals self-select light patterns in the real world⁴⁸. Specifically, an indoor light level of 50 lux was used for both work hours and hours of wakefulness before sunrise or after sunset^49–51. A smooth light profile with a maximum of 300 lux was used for non-work hours (including potential commuting hours) between sunrise and sunset:

$$I=150\tanh \left(c\left(t-s_1\right)\right)-150\tanh \left(c\left(t-s_2\right)\right),$$ (4)

where $t$ is clock time (time in hours, modulo 24, ranging from 0 to 24). The parameters $s_1=7.5\mathrm{h},s_2=16.5\mathrm{h}$, and $c=0.6{\mathrm{h}}^{-1}$ were chosen to generate an approximately 12-hour photoperiod centered at noon. Light levels were set to zero lux during any times the individual was asleep; therefore, light levels rise and fall abruptly at sleep offset and onset, respectively.

Prior to simulating the work schedules, the model was entrained to a permanent day-shift schedule for 6 months with two non-work days per week and sleep opportunities from 22:30 to 5:30 on the five work days per week; this was to ensure entrainment of the circadian pacemaker and equilibration of the long timescale sleep homeostatic component to the schedule. This schedule was motivated by the pre-study conditions in Lockley et al.¹¹ and corresponds to the sleep times used in a previous simulation paper²⁶. To obtain PVT thresholds corresponding to moderate impairment and high impairment, a simulation was performed in which the model was held awake from a wake time of 5:30 for 16 hours and 20 hours, respectively, at the end of the baseline schedule; note that the end of these continuous wake durations is during the circadian “night” and therefore during circadian misalignment. The threshold for moderate impairment was defined as the number of PVT lapses predicted by the model at 16–20 hours of continuous wakefulness: 8.7–15.2 lapses. Similarly, the threshold for high impairment was defined as the number of PVT lapses predicted by the model at ≥20 hours of continuous wakefulness: >15.2 lapses.

Analysis

As primary outcomes for each schedule, we computed the average number of hours per week (i.e., per 168 hours) during working hours corresponding to: (i) moderate impairment; and (ii) high impairment. The EDWR and RCWR conditions were first compared for all weeks of data in all participant schedules. Because the impairment time distributions were highly skewed, a nonparametric comparison was used (Wilcoxon rank sum test).

To measure acute vs. chronic effects of each schedule, number of hours of moderate impairment and high impairment were computed for weeks 1+2 and for weeks 3+4 on each schedule, using all available data within the time range. A mixed-effects regression model was fit to the primary outcomes, including schedule and week block as fixed effects and participants as random effects. The week 1–2 block in the EDWR condition (i.e., first time point in the non-intervention condition) was used as the baseline for contrasts. An interaction between schedule and week block was included. Normality of residuals was checked with Q-Q plots. Mixed-effects models were implemented in MATLAB (R2015a, Mathworks, Natick MA, USA) using the function fitlme. Absolute effect sizes and their 95% confidence intervals were reported.

To examine temporal patterns in performance, the mean of predicted PVT lapses were computed for each schedule in 1-h clock-time bins across the schedule. Two methods of averaging were used for each schedule: (i) using all hours of wakefulness, and (ii) using all work hours. To compare between schedule conditions, these temporal patterns were averaged across schedules within each condition, and 1-h timepoint-wise comparisons between conditions were made using t-tests after running an ANOVA to confirm differences between conditions. Differences were considered significant if p<0.002 (with Bonferroni correction for multiple comparisons).

As a secondary analysis, we investigated whether performance differences within the EDWR or RCWR conditions could be explained by total average sleep time. For each individual schedule, we computed the average percentage of time spent asleep in 1-h clock-time bins from start to end of schedule, so that estimates would not be biased by clock timing of missing data. The average total sleep time for an individual schedule was computed by averaging across clock-time bins. Spearman correlations were computed between average total sleep time and hours per week in a predicted moderate or highly impaired state.

As a further secondary analysis, we investigated whether performance differences within the EDWR or RCWR conditions could be explained by the frequency of short between-shift intervals (defined as intervals <10 h in duration), since they could affect sleep timing and duration. Specifically, we computed the average number of short between-shift intervals per week (168 hours). Spearman correlations were computed between the frequency of short between-shift intervals and hours per week in a moderate or highly impaired state.

Results

Examples of predicted amount of time spent in a moderately or highly impaired state on EDWR and RCWR are shown in Figure 1. On the EDWR, there was a range of 5.0–47.1 hours per week spent in a moderately impaired state (median=15.5 h), while there was a range of 1.4–32.2 hours per week (median 13.5 h) spent in a highly impaired state. On the RCWR, there was a range of 3.6–37.1 hours per week (median=10.8 h) spent in a moderately impaired state, while there was a range of 0–37.6 hours per week (median 12.0 h) spent in a highly impaired state (Figure 2). At this schedule level, time spent in a moderately impaired state was significantly higher for EDWR compared to RCWR (p=0.02; effect size = 0.2). Time spent in a highly impaired state was not significantly different between schedules (p=0.11).

Figure 1:

Double raster plots for examples of (A) the Extended Duration Work Roster (EDWR) and (B) the Rapidly Cycling Work Roster (RCWR). Time intervals corresponding to sleep are marked by black bars. Time intervals corresponding to work are marked by white, yellow, and red bars, which represent times of predicted low, moderate, and high performance impairment during work, respectively.

Figure 2:

Predicted hours per week during work hours spent under (A) moderate impairment (equivalent to 16–20 hours of continuous wakefulness); and (B) high impairment (equivalent to >20 hours of continuous wakefulness). Distributions for the Extended Duration Work Roster (EDWR) and the Rapidly Cycling Work Roster (RCWR) are shown as box-and-whisker plots. Individual data points (169 schedules total; 95 EDWRs and 74 RCWRs) are plotted as circles, with x-axis position jittered to reduce overlap of points. Each individual point in each panel is one simulated schedule.

Mixed-effects modeling showed significant main effects of the week block and a significant interaction of condition with week block. For the main effect of week block, we found that compared to the week 1–2 block, time spent in a moderately impaired state per week was increased by 2.6 hours (p=0.001, 95% CI: 1.0–4.2 hours) and time spent in a highly impaired state was increased by 1.7 hours (p=0.001, 95% CI: 0.7–2.8 hours) in the week 3–4 block. For the interaction of condition with week block, we found that compared to RCWR, time spent in a moderately impaired state for EDWR was further increased from weeks 1–2 to weeks 3–4 by 3.4 hours (p=0.003, 95% CI: 1.2–5.7 hours), whereas time spent in a highly impaired state had no significant interaction (0.6 hours, p=0.46, 95% CI: −0.9–2.0 hours). There was no main effect of condition, meaning time spent in a moderately or highly impaired state in the week 1–2 block did not significantly differ between conditions.

Differences in predicted performance with time of day were observed between the schedules (Figure 3). For the analysis of performance across all hours of wakefulness (Figure 3A), there was a main effect of condition (p<0.0001) and time (p<0.0001). In pairwise comparisons, performance was significantly worse on the EDWR than the RCWR in all clock time bins from 21:00 to 04:00. Similarly, for the analysis of performance across working hours (Figure 3B), there was a main effect of condition (p<0.0001) and time (p<0.0001). In pairwise comparisons, performance was significantly worse on the EDWR than the RCWR in all clock time bins from 19:00 to 05:00, whereas performance was significantly better on the EDWR than the RCWR in all clock time bins from 08:00 to 11:00.

Figure 3:

Mean predicted attentional failures as measured by PVT lapses for Extended Duration Work Roster (EDWR) and Rapidly Cycling Work Roster (RCWR) schedule conditions, plotted by time of day averaged across all data sets (169 schedules total; 95 EDWRs and 74 RCWRs) in 1-hour clock time bins. Data are plotted for (A) all hours of wakefulness, and (B) working hours, and plotted at the midpoint of each clock hour. Error bars show mean ± SEM. Significance is shown for timepoint-wise comparisons: *p<0.002, **p<0.0001.

Short between-shift intervals were significantly less common on the EDWR than the RCWR (0.3% vs.8.4%, p<0.0001)⁴⁷. Within the EDWR condition, the frequency of short between-shift intervals was not significantly associated with predicted number of hours in a moderately impaired state (r=−0.15, p=0.07) or highly impaired state (r=−0.02, p=0.84). Within the RCWR condition, however, the frequency of short between-shift intervals was significantly associated with predicted number of hours in a moderately impaired state (r=0.39, p<0.000001) but not with predicted number of hours in a highly impaired state (r=−0.01, p=0.92).

Predicted performance was highly correlated with sleep duration (Figure 4). On average, total subjective sleep time (from daily logs) was significantly lower on the EDWR schedules compared with the RCWR schedules (6.7 ± 0.7 h vs. 7.0 ± 0.7 h; p=0.002, effect size = 0.20). Spearman correlations of total subjective sleep time with work hours spent in a moderately impaired state were significant for the EDWR (r=−0.85, p<0.001) and RCWR (r=−0.81, p<0.001). Spearman correlations of total subjective sleep time with work hours spent in a highly impaired state were also significant for both the EDWR (r=−0.63, p<0.001) and RCWR (r=−0.70, p<0.001).

Figure 4:

Associations of average total subjective sleep time across the entire schedule for all 169 schedules (95 EDWRs and 74 RCWRs) with predicted hours per week spent in (A) moderate impairment during work (equivalent to 16–20 hours of continuous wakefulness); and (B) high impairment during work (equivalent to >20 hours of continuous wakefulness). Each individual point in each panel is one simulated schedule.

Discussion

We used a mathematical model to compare traditional EDWR schedules with RCWR schedules (that were designed to ameliorate impacts of working hours on sleep and neurobehavioral performance) using predicted neurobehavioral performance during 169 resident physicians’ sleep and work schedules. This is the first example of linking three key components for real-world performance prediction in a single study: (i) a dynamic model of the human circadian pacemaker and its response to light/dark and sleep/wake patterns; (ii) a model of effects of chronic (cumulative) sleep restriction on performance; and (iii) actual schedules of self-selected sleep/wake timing with associated realistic light/dark patterns. Our findings indicate that the RCWR intervention is partially effective in reducing the predicted neurobehavioral performance impairments generated by the traditional EDWR. Specifically, the RCWR intervention was found to be most beneficial during the circadian (biological) night and least beneficial in late morning.

A key strength of this analysis is the use of real sleep and work patterns, rather than idealized or planned schedules. Within both the EDWR and RCWR schedules, we observed considerable interindividual and intra-individual variability in both sleep and work patterns. This variability is representative of real-world differences in compliance, adherence, work/social constraints, and physiology, as well as schedule implementation⁴⁷. Using the model across these datasets, we were able to identify two concrete areas of future improvement for RCWR schedules. First, we found a specific window of vulnerability in performance during work hours from 08:00 to 11:00, which is not predicted to be present on EDWR schedules. Second, we identified that short between-shift intervals are predictive of a greater number of hours working in a moderately impaired state specifically for RCWR schedules. This prediction suggests that carefully controlling the duration of between-shift breaks could enhance efficacy of the RCWR intervention.

Our work predicts chronic deterioration of performance on both RCWR and EDWR schedules. Compared to EDWR, the rate of deterioration is predicted to be lower by 57% on the RCWR. As experimental studies have demonstrated, the effects of chronic sleep restriction can accumulate over weeks^16,38,52,53. Moreover, if chronic sleep restriction is alternated with recovery breaks that are of insufficient duration to achieve full recovery to baseline, the rate of cumulative decline in neurobehavioral performance may not be slowed by the recovery intervals¹⁹. Studies of neurobehavioral performance in resident physicians to date have all been of less than a month in duration^12,54. Our findings suggest the need for longitudinal studies, across months or years, to assess the long-term impacts of resident physician schedules on sleep and neurobehavioral performance.

We note that the model used in this study predicts objective performance on the PVT, rather than subjective measures of alertness. Subjective alertness, which is used by individuals to decide if they need a countermeasure (e.g., nap or caffeine) and/or if they can safely perform (e.g., driving home) is not well correlated with objective performance when individuals are exposed to chronic sleep restriction^{19,38,53,55,56}. Adding subjective alertness to the model could improve its use in education about the dangers of insufficient sleep, including its effects on objective performance.

There are possible limitations of the method presented here. First, all components of the model have been trained on healthy young adults with intermediate chronotypes (i.e., excluding individuals whose preferred sleep timing is extremely early or extremely late). Predictions may therefore not be representative for individuals of other ages; with sleep, circadian, or other disorders; and/or extreme chronotypes. In the future, the model could be used to investigate sensitivity of the predictions to physiological parameters that determine chronotype, as has been done previously for diurnal schedules^44,48, since different individuals may respond differently to the same work or sleep schedule^57,58. Second, this model does not yet incorporate the effects of caffeine or other pharmaceuticals on sleep and performance. Models of caffeine’s effects on sleep and neurobehavioral performance have been developed^59–61 and should be incorporated in future. Models of the effects of prescription medications need to be developed. Third, the model predictions generated from these real work patterns have not been validated against data. The models, however, have previously been validated on data from other experimental studies of chronic sleep restriction and circadian misalignment of sleep/wake and light/dark patterns³⁷. Validation on experimental data from RCWR and EDWR schedules is an important next step to confirm whether some of the differences observed between RCWR and EDWR, such as the worse performance predicted from 08:00 to 11:00 on the RCWR schedule, are observed during actual work shifts. Finally, the model predictions do not currently account for effects of sleep inertia on neurobehavioral performance^62,63. Notably, many of the participants on EDWR schedules took naps during the extended duration shifts, followed by immediate return to work. The model simulates the restorative properties of napping, but not the sleep inertia, which may be impactful even for short naps²⁵.

In previous work with the chronic sleep restriction component of the model, the model was also used to generate accurate predictions of sleep/wake patterns³⁷. Here, we chose instead to use the reported sleep/wake patterns for each schedule as input to the model. This selection was made for two reasons: (i) the model has only been validated for sleep/wake predictions on daytime work schedules, and sleep during rotating shiftwork will be more difficult to predict; and (ii) to ensure that performance predictions are not systematically inaccurate due to any potential inaccuracies in the model’s predictions of sleep/wake timing. In future work, the model’s sleep/wake predictions could be trained against this and other datasets. This would allow model-based prospective assessment of candidate schedules in the absence of any worker data to determine which schedules allow the most opportunity for sleep with the least amount of time spent in a moderately or highly impaired state.

Modeling is an essential component of effective occupational risk assessment, yet the use of models in medical settings remains very limited^26,64. The modeling approach presented here could be employed in an active role in future medical residents or other schedule design and optimization (e.g., transportation [airline pilots, train conductors, ship captains, truck drivers], military/security, nuclear power plant monitoring) for individuals and teams. Although there are many factors – both fiscal and logistical – that constrain the design of resident physician schedules, our findings underscore the importance of considering the effects of sleep loss and adverse circadian phase, especially given the high cost of failure to both patient safety and physician health in the medical setting or operator and public safety in other settings.

Public Health Relevance.

This work highlights the utility of mathematical models for evaluating the predicted performance profile of work schedules in safety-sensitive occupations that are susceptible to chronic sleep restriction and circadian misalignment, including resident physicians; individuals working in transportation, military/security, and industry; and others.