We conducted a nationwide prospective cohort study of resident physicians in the US for eight academic years (2002-07 and 2014-17). All study procedures were approved by the Partners Human Research Committee and a certificate of confidentiality was issued by the US Centers for Disease Control and Prevention.
Data collection
With assistance from the Association of American Medical Colleges, US medical school graduates and individuals who matched to a US residency programme from 2002 to 2006 were invited by email to participate in an online study. Similarly, from 2014 to 2016, all medical school graduates who completed an application through the Electronic Residency Application Service were invited by email to participate in the study. The invitation directed potential participants to the study website where information about the study, excluding the study hypotheses, was provided.
Those individuals who chose to enrol in the study completed an electronic consent form and provided an email address where they wanted to receive their monthly surveys. Throughout the course of the study, they had the opportunity to update their email address (eg, change from a school email address to a hospital or personal email address). At the end of each academic year, participants were invited to continue their participation into their next year of residency. On enrolment, each participant was given a unique ID enabling the linking of the surveys completed by that individual throughout the course of their participation. After their first year of participation in the study, resident physicians were invited to continue their participation as more senior resident physicians. Nominal incentives (eg, $30 (£25.05; €28.39) for completing five monthly surveys) were provided as well as randomly drawn cash prizes.
In June of each year, unique individual links were sent via email to resident physicians who consented to participate in the study. The baseline survey collected information, on personal characteristics, including age, gender, height, weight, medical history, and specialty programme. Monthly reports collected work hour information, including total hours of work, frequency of shifts of extended duration (≥24 hours), hours engaged in patient care, and additional work related to their residency programme, using previously validated methodology.3 Hours of sleep at work and away from work were reported. In a separate section of the report, participants reported important medical errors in the past month and any patient outcomes resulting from errors. Errors resulting in patient harm were considered preventable adverse events. Participants also reported on the frequency of adverse health and safety outcomes, including motor vehicle crashes and near miss crashes. The instrument used to collect information on motor vehicle crashes has been previously validated.3 We also collected information on the frequency of occupational exposures to potentially contaminated blood or other bodily fluid and queried how many times participants nodded off or fell asleep (attentional failures) during inappropriate times (eg, surgery; rounds with attending physicians; while talking to or examining patients; and in lectures, seminars, and grand rounds). See online supplemental materials 1 for survey questions.
Statistical analysis
Given that the major components of the ACGME work hour guidelines did not differ substantially for PGY2+ resident physicians in our 2002-07 and 2014-17 study cohorts (ie, 80-88 hour work week limit, 28-30 consecutive work hour limit, including time for transitions; minimum four days off per month), we pooled the responses for our analyses, although we controlled for cohort and other potentially confounding factors in multivariable analyses. We excluded months when participants reported ≥14 work-free days (vacation months), and when work hour information was missing or reported to exceed 168 hours of work per week. Weekly work hours were calculated as the sum of the number of hours spent physically awake in the hospital, classes, or workplace, plus the number of hours asleep in the hospital.
We examined the association among weekly work hours, shifts of extended duration, and adverse health and safety outcomes. The incidence of adverse health and safety outcomes were calculated. We tested the significance of the calculated incidence rate ratios using likelihood ratio tests in log-linear models. Sensitivity analyses used Pearson and deviance based, scaled Poisson models that accounted for overdispersion, obtaining similar results. We dichotomised rare outcomes to reflect the presence or absence of at least one outcome during the month, and estimated the risk of each outcome using generalised linear models with a binomial distribution and log-link function. Basic models for clinical outcomes were adjusted for reported hours of patient care that month. We identified potentially confounding variables a priori based on relevance to the research question and biological plausibility. Fully adjusted multivariable models controlled for age, gender, specialty programme, cohort, and an age imputation indicator variable. Analyses stratified by cohort are presented in online supplemental tables S1A and S1B.
For clinical outcomes (medical errors, preventable adverse events, medical errors resulting in patient death, occupational exposures, and percutaneous injuries), we also controlled for the hours of patient care reported that month. Mixed effects models were used to examine shifts of extended duration or weekly work hours as the independent variable of interest, with a random intercept for participant to control for the dependence between repeated measures. We conducted analyses with weekly work hours as a continuous variable and separately as a categorical variable with hours grouped into categories to inform policy decisions. Missing data for age (total 205 participants, 1561 person months (4% of months)) were imputed with the median age of the participant cohort, and a binary variable was created to indicate that age was imputed for that participant’s observations (age imputation indicator variable).4 We did sensitivity analyses with and without accounting for missing data to guide the final models presented (online supplemental tables S4A-C, figure S1).
Categorical analyses used ≤48 hours as a referent because it is the European Working Time Directive's limit. We chose additional, reasonable cut-off points (60, 70, and 80 h/week) that had sufficiently distributed data and that could be analysed and reasonably translated to meaningful policy. Further, because the ACGME limit on work weeks is 80 hours, extended routinely to 88 hours, we dichotomised the analysis (>80 v ≤80 hours), to evaluate the risks associated with working above the limit.
Motor vehicle crash models were limited to participants who reported a valid driver’s license and who also reported driving to or from work. We limited medical error and occupational exposure models to months in which participants reported hours in patient care. SAS (version 9.4, SAS Institute, Cary, NC) was used for statistical analysis. All tests were two sided and P<0.05 was considered significant.
Patient and public involvement
We have worked extensively with public stakeholders, patient advocates, and resident physician advocacy organisations in our work on the effects of sleep deficiency on patient safety. We plan to widely disseminate the results to relevant patient and public communities, including organisations such as the Committee for Interns and Residents, Patient Safety Action Network, Agency for Healthcare Research and Quality, ACGME, National Academy of Medicine, and Royal College of Physicians. We will also post the results to the study website.