Data Abstraction

Of the 343 papers reviewed, 42 studies described relevant interventions, and 11 of those evaluated patient outcomes. We eliminated 2 studies because they relied solely on survey reports of errors (1 resident self-report [14] and 1 report of nurses' perceptions of resident errors [15]). We eliminated 2 other papers because they studied patient satisfaction, not safety (16, 17). Thus, we included 7 studies in this review. Two authors abstracted the data from each included article, and 1 author reviewed all studies. We used a standardized abstraction form that included the following information: number of participants, presence or absence of a control group, study design, outcomes, and methodologic concerns. We resolved all disagreements by discussion and consensus. When necessary, we attempted to contact the authors of the studies to provide additional information.

Role of the Funding Source

The Ann Arbor Veterans Affairs Medical Center/University of Michigan's Patient Safety Enhancement Program had no role in the design, conduct, or reporting of the study or in the decision to submit the manuscript for publication.

Data Synthesis

We present the results of the 7 intervention studies that assessed patient safety outcomes. The Appendix Table shows the study design, strengths, and weaknesses of each study. No study was a clinical trial. The Table summarizes the studies' diverse interventions. The outcomes also varied considerably from study to study (Table). To clarify our terminology, we use "interns" to refer to physicians in their first year of postgraduate training and "residents" for those in at least their second year of postgraduate training.

Daigler and colleagues (18) conducted a retrospective cohort study to assess the effect of schedule changes for interns in a pediatrics program that would result in compliance with New York code 405 (the New York state law that limited resident work hours in 1989 [25]). In the preintervention period (1987–1988), residents worked approximately 100 hours per week and took overnight call every third or fourth night. In the postintervention period (1988–1989), the interns were scheduled to work less than 70 hours per week. Each week had one 24-hour day, one 16-hour day, one 12-hour day, three 9-hour days, and one day off. The intensive care units continued to have call every third night. The authors reported no difference in mortality, morbidity, or unexpected intensive care unit transfers after the schedule changes occurred. However, this study is flawed because the methods section contains very limited information about the patient sample, data collection, and data analysis.

Gottlieb and colleagues (19) used a prospective pre–post design to evaluate the effects of schedule changes in an internal medicine program at a Veterans Affairs hospital. These changes consisted of shifting from a call system with interns on call every fourth night and residents on call every eighth night to a system that incorporated a night- float system. The teams in the new system had 7-day cycles with 1 long call day, 3 short call days, and 3 noncall days during each cycle. Short call also occurred on the weekends. Therefore, interns were on call overnight every seventh night while residents were on call overnight every 14th night. Night-float teams admitted all patients after 10 p.m. on Sunday to Thursday. The housestaff reported sleeping substantially more after the intervention (19). The patient-related outcomes studied were medication errors, fevers, deaths, and readmissions. Resource utilization was also studied in the form of length of stay and the numbers of laboratory tests, consultations, and radiographs ordered. Statistically significantly fewer medication errors occurred after the schedule changed. Incidence of fever, death, or readmission did not differ. Length of stay and number of laboratory tests ordered both decreased after the intervention, and the numbers of radiographs or consultations ordered did not differ. Although this nonrandomized study included an initial comparison of baseline characteristics of the patients, the authors did not adjust for them in their analyses.

Howard and colleagues (20) performed a retrospective study by using a statewide database to assess the effect of New York code 405 (25) on mortality. They used a before-and-after approach to look for a mortality difference in patients with congestive heart failure, acute myocardial infarction, or pneumonia who were admitted to teaching hospitals in 1988 (before) and 1991 (after). They used patients admitted to nonteaching hospitals as controls to account for temporal trends. They found that mortality decreased in all 3 diagnostic groups (and in the combined group) from 1988 to 1991, but the decrease was equal between the teaching and nonteaching hospitals; this finding suggests that New York code 405 (25) could not be responsible for the decrease (20).

Laine and colleagues (21) conducted a retrospective study of general internal medicine patients admitted before and after an intervention. The original resident schedule was an every-fourth-night call schedule. The intervention was to continue the every-fourth-night call cycle, but the interns left after the shift and their patients were cross-covered by another intern. Residents went home at 11 p.m. on call nights, and a night-float resident served as back-up for the intern who was on call. The outcomes evaluated were mortality, procedure delay, diagnostic test delay, and complication rates. They found that mortality and procedure delays did not differ before and after the intervention, but overall complication rates and diagnostic test delays were statistically significantly more common after the intervention. This nonrandomized trial adjusted for some baseline characteristics in the analyses. The small number of individual complications (such as drug reactions, line infections, and electrolyte abnormalities) makes it difficult to draw conclusions about specific risks.

Lofgren and colleagues (22) performed a retrospective cohort study of patients admitted to a general internal medicine service at a Veterans Affairs hospital. They evaluated the differences between 2 types of admission teams: the primary intern working with a cross-cover resident or the primary intern and primary resident working together. Mortality did not differ between the 2 groups of patients. Resource utilization in the form of length of stay, consultations, procedures, and radiographs did not differ between the groups. However, the number of laboratory tests ordered per patient was higher in the group with the cross-covering residents (22). This study's strength was the measurement of baseline characteristics between the groups, although systolic blood pressure was higher in the group with the cross-covering resident. Its weakness is that not all analyses adjusted for baseline factors. In addition, 24% of the sample was not included in the final analysis, and it is not clear whether the dropout rates differed between the 2 groups.

Mann and Danz (23) used a prospective cohort study design to evaluate "discordant" cases occurring during night coverage in a radiology program. Discordant cases were those read initially by the on-call radiology resident and then changed by the attending physician the next day. The intervention in this study was the addition of a "night stalker," a night-float resident who assisted the on-call residents 4 nights per week. The amount of sleep by the regular on-call residents increased (although this difference was not tested for statistical significance). The authors reported that the night stalker had fewer discordant cases per shift than the on-call residents. No difference in patient morbidity was detected (23). A weakness of this study is that it did not compare or adjust for baseline patient characteristics.

Petersen and colleagues (24) conducted a prospective case–control study of adverse events in patients admitted to internal medicine services. They compared 3 simultaneous systems of coverage: a primary team, "cross-coverage" by another team's intern, and a night-float system. They reported that non–primary team coverage (cross-coverage and night-float system together) conferred a statistically significantly greater risk for adverse events than primary team coverage. When cross-coverage and night-float systems were considered separately, only cross-coverage had an increased risk (24). This study's strengths include the prospective nature, the matching process, and the measurement of baseline characteristics. Its weakness is that the cases and controls differed on several important baseline characteristics (for example, Acute Physiology and Chronic Health Evaluation [APACHE] score). However, these differences were accounted for in the multivariable model.

Discussion

Top Methods Discussion Author & Article Info References

Defining the best practices for improving patient safety is controversial and complex (4, 26). In the 2001 report "Making Health Care Safer: A Critical Analysis of Patient Safety Practices" (26), the authors conclude that evidence is insufficient to estimate the effect of an intervention to limit physician work hours. However, on the basis of the literature (largely outside of medicine) of the effect of sleep deprivation on performance (27), the authors advocate efforts to reduce both fatigue and sleepiness (26).

In the medical profession, some studies have evaluated the link between performance and sleep deprivation (28, 29). For example, residents perform worse on laparoscopic surgery simulators after a sleepless night than after a normal night (30). The studies of the effects of sleep deprivation on performance have focused mainly on tasks, and task performance tends to degrade with sleep deprivation (31-33). The central question, however, is more complex than merely the performance of tasks. Who provides safer care to patients whose health is in flux? A well-rested resident who knows the patient primarily through a report or a fatigued resident who has firsthand knowledge through performing the admitting history and physical examination? The relationship among continuity of care, fatigued providers, and patient safety is complex (26). Efforts to reduce fatigue and sleep deprivation among residents often result in increased discontinuity of patient care, and, as noted, the link between discontinuity and patient safety may also be harmful (26, 34). This balance between promoting and preserving continuity while minimizing physician fatigue must be central to the ongoing resident work hour reforms (35).

Our results must be interpreted cautiously. First, these studies are of variable quality, with different strengths and weaknesses, and none was a clinical trial. Second, the interventions used differ among studies; therefore, generalizations are difficult to make. Third, the results of individual studies are conflicting. In addition, publication bias may be present in that negative studies may not have been published and would therefore not be represented in the literature.

Although 3 of these studies (20, 21, 24) adjusted for differences in their analyses, unmeasured confounding is still a potential problem of nonrandomized studies (36). One of these studies reported that patients were more likely to have a complication or a test delay after schedule changes occurred (21). In the second study, the only case–control study, the authors used conditional logistic regression to account for confounding on the basis of the matching process, and in the multivariable model, potentially preventable adverse events were significantly associated with coverage by a nonprimary team (24). The third study showed a decrease in mortality at teaching hospitals after adherence to New York code 405 (25) that was mirrored at nonteaching hospitals (20). The fact that 2 of 3 multivariable models demonstrated some risk associated with different interventions to decrease resident work hours may be important, but the models are limited by their inability to account for unmeasured confounding; this is probably the most important limitation of these well-done, nonrandomized studies.

The studies in our review analyzed different interventions to limit resident work hours, although 4 of the 7 studies included a night-float system. The responsibilities of the night-float resident also differed among the interventions. These differences may explain some of the conflicting results. Any intervention to limit work hours may decrease continuity of care. On the other hand, perhaps the continuity between admission and the remainder of care is more important than continuity from hour to hour. In 1 study, the authors contend that their intervention (which included a night-float system) actually improved continuity for admissions (19). Before the interventions were implemented, residents admitted with their interns on only half of the interns' long call nights. In the new system, the interns and residents on a given team admitted together more often. Of interest, this study showed fewer medication errors after the schedule changed (19). The increase in continuity may have contributed to fewer medication errors. Medication errors (a task-based activity) may also be more sensitive to fatigue than are complications and test delays, both of which may be more sensitive to decreased continuity.

Complication rates increased and diagnostic test delays occurred in a study that replaced the resident on call at 11 p.m. with a night-float resident, thereby decreasing continuity (at the resident level) for the patients admitted late at night (21). The primary team may not get to know their patients at the same level of detail if the resident does not participate in the actual admission, even if the intern does participate. This seems to be corroborated by the finding that more laboratory tests per patient were ordered when a cross-covering resident admitted patients with an intern when compared with the primary resident–primary intern admission team (22). However, length of stay, mortality, consultations, and radiographs did not differ between the groups in that study (22).

We encourage the application of evidence-based medicine in future graduate medical education policy changes. To that end, we offer the following suggestions for a research agenda. First, research funding must be invested in carefully evaluating outcomes in graduate medical education. Only 1 study in our review reported that external funds were used for the project (24). These policy changes affect the many Americans who seek medical care at teaching hospitals. The effects of such decisions should be studied at high levels of quality, which is possible only through adequate grant support.

Second, patient safety outcomes should be evaluated. While resident performance on simulation exercises has been measured (30-32, 37), we believe the most compelling evidence would be measured in the context of caring for actual patients. Third, investigations should be coordinated across several institutions. This is probably the only way to achieve meaningful numbers and external validity. The interventions should be as similar as possible to allow for pooling of data. Another alternative is to use statewide databases, as 1 of the studies in our review did (20). Fourth, if possible, trials should be randomized. When clinical trials are not possible, measuring and adjusting for potential confounding variables should be routinely performed. Finally, consideration of discontinuity between admission and the remainder of the hospitalization separately versus discontinuity in the hour-to-hour care of patients may be useful.

Safe patient care is our profession's most important priority and must remain in the forefront as efforts to reform resident work hours continue. As this systematic review illustrates, only a small body of imperfect literature is available on the outcomes of past interventions to decrease housestaff work hours, fatigue, and sleep deprivation. These inconclusive findings are important. The general public believes that limiting hours will improve patient safety (38, 39); however, our review suggests that the effect of past efforts has been unclear. In addition, our review further clarifies the methodologic challenges in determining the effect of work hours reform. To be confident of changes in mortality rates, for example, much larger, multi-institutional studies with careful measurement for possible confounders are required. On the basis of the inconclusive results of the studies performed to date, we advocate some caution as U.S. residency training programs conform to the ACGME guidelines. While the reduction of resident work hours will probably be permanent, defining, maintaining, and studying the interventions that provide the appropriate balance between continuity of care and physician fatigue should be a central consideration to ensure that our current reform efforts improve patient care.