Article Text
PDF
Abstract
Background Understanding factors that drive admissions is critical to containing cost and optimising hospital operations. We hypothesised that, due to multiple factors, emergency physicians would be more likely to admit a patient seen later in their shift.
Methods Retrospective study examining all patient visits at a large academic hospital from July 2010 to July 2016. Patients with missing data (n=191) were excluded. 294 031 emergency department (ED) visits were included in the final analysis. The exposure of interest was the time during the shift at which a patient was first evaluated by the clinician, and outcome was hospital admission. We used a generalised estimating equation with physician as the clustering level to adjust for patient age, gender, Emergency Severity Index (ESI, 1=most severe illness, 5=least severe illness) and 24 hours clock time. We also conducted a stratified analysis by three ESI categories.
Results From the 294 031 ED visits, 5977 were seen in the last hour of the shift. Of patients seen in the last shift hour, 43% were admitted versus 39% seen at any other time during the shift. There was a significant association between being evaluated in the last hour (RR 1.03, 95% CI 1.01 to 1.06) and last quarter (RR 1.02, 1.01 to 1.03) of shift and the likelihood of admission. Patients with an ESI Score of 4–5 saw the largest effect sizes (RR 1.62, 0.996–2.635 for last hour and RR 1.24, 0.996–1.535 for last quarter) but these were not statistically significant. Additionally, there was a trend towards increased likelihood of admission later in shift; the relative risk of admission was 1.04 in hour 6, (1.02–1.05), 1.03 in hour 7 (1.01–1.05), 1.04 in hour 8 (1.01–1.06) and 1.06 in hour 9 (1.013–1.101).
Conclusions There is a small but significant association between a patient being evaluated later in an emergency physician’s shift and their likelihood of being admitted to the hospital.
- cost-effectiveness
- emergency department
- healthcare quality improvement
- implementation science
- quality improvement
Statistics from Altmetric.com
- cost-effectiveness
- emergency department
- healthcare quality improvement
- implementation science
- quality improvement
Introduction
Hospital admission is a major driver of healthcare costs, and the emergency department (ED) is a major source of such admissions.1 While ED rates of admission are decreasing overall, the total number of visits to EDs has increased dramatically over the last decade; as such, the total number of admissions is increasing.2 To better understand and contain rising healthcare costs, understanding the upstream factors that drive hospital admissions is crucial.
As part of our efforts to standardise decisions regarding admission, an important area of study remains incomplete: What factors drive variation in provider behaviour? Prior research suggests that both provider-level and systems-level factors are at play. Studies of all admissions within a health system,3 as well as for specific conditions—chest pain,4 pneumonia5 and skin infections6—demonstrate significant variation in provider behaviour. Authors hypothesise that such differences may be due to provider education and training, differential risk tolerance, willingness to use clinical decision rules and fear of malpractice. However, systems-level factors also influence admission decisions. Both hospital size7 8 as well as ED crowding9 10 have been shown to increase rates of admission.
One additional important consideration is the effect of time of shift on provider behaviour. As shifts progress, providers may acquire more tasks and patients competing for cognitive resources. At the end of a shift, patients who need additional diagnosis or treatment may be signed out by the departing provider to an oncoming provider who assumes their care. At sign-out, there may be significant pressure on the departing provider to have a clearly formulated plan for diagnosis, treatment and disposition of the patient.11 Due to multiple factors, providers may be more likely to admit a patient later in the shift.
Our goal was to examine the relationship between the time during a provider’s shift at which the patient is evaluated by the emergency provider and the likelihood of admission to the hospital. We hypothesised that there would be a significant relationship between later hours of a provider’s shift and patients’ likelihood of admission to the hospital.
Methods
Research integrity and human subjects protection
This observational cohort study is reported in accordance with the STrengthening the Reporting of OBservational studies in Epidemiology statement.12
Study description
This was a retrospective observational cohort study of data from the ED at a single academic centre in Boston, Massachusetts, USA, with approximately 55 000 annual visits. The data are drawn from our emergency medicine operations database which collects information about all visits to our ED. All patients who visited our ED from July 2010 to July 2016 were included for preliminary analysis; patients with missing data (n=191) were excluded. In the final analysis 294 031 ED visits were included. Though many patient visits are shared with residents, we considered the attending provider as ultimately responsible for admissions, and the times of attending provider shifts mirror those of resident shifts.
Exposures and outcomes
The primary outcome was the binary outcome of admission to the hospital, meaning that the patient was moved from the ED to an inpatient bed. We did not differentiate between inpatient observation status versus full inpatient admission. We did not consider ED observation status as an admission.
The primary exposure of interest was the time during the shift at which a patient was first evaluated by the clinician, which we termed ‘shift hour’, represented as an integer. Patient evaluation begins when a provider electronically ‘claims’ a chart in the medical record, and the culture in our institution is that providers immediately evaluate patients claimed. We thus used the timestamp from the physician ‘claiming’ the chart to determine the primary exposure.
To determine what hour of the provider’s shift was associated with seeing the patient, we divided the provider’s shift into bins of 1 hour length (eg, we used the hour as the unit of time). We determined the start time and length of each provider’s shift, and by evaluating the time when the chart was electronically claimed (the ‘door-to-doctor time’ as logged in our electronic record), we could place the patient into a shift hour bin for that provider on a specific visit (hour 1, hour 2, etc). Only full hours were used; for example, if a shift began at 14:00, and the patient was seen at 14:47, this patient was described as being seen in ‘hour 1’. Of note, this is a different concept than the time the disposition decision was made. Our group has previously found that admission decisions are often delayed by several hours from their initial presentation; the timing is primarily related to diagnosis and acuity and is therefore less likely to reflect cognitive fatigue, pressure to establish a ‘clean’ sign-out and other factors.13
Each entry in our data set includes details about a patient’s encounter with specific providers, and the primary exposure is calculated for each individual provider. We excluded sign-out patients (those patients whose care had been transitioned to another provider). When patients are signed out, a new entry is generated for that patient in our database; each entry is assigned a binary variable (sign-out yes/no=1/0). We included only patients who had not been signed out. Of note, these entries are separate from patients with missing data.
We classified our primary exposure, shift hour, in three ways:1 the last hour of shift compared with all previous hours of that shift (binary outcome);2 the last quarter of shift, compared with the first three quarters (binary outcome); and3 the shift hour as an ordinal integer ranging from 1 to 9 depending on the length of the provider’s shift, compared with the first hour. This was done to gain thorough insight about the nature of the relationship between shift time and provider behaviour.
Patients are picked up ad libitum from an electronic rack of patient charts by the resident, seen by the resident and then staffed with the attending. We elected to model the primary exposure by attending, as the attending ultimately controls the disposition decision. Attending and resident shifts directly align (eg, if an attending works 2:00–11:00, a resident works the same shift); there are no dyssynchronous attending/resident shifts. There are no mid-level providers (nurse practitioners or physician assistants) at our hospital. With respect to claiming charts, residents are expected to claim the next chart in the queue. ‘Cherry picking’ (that is, selectively choosing charts which may be less cognitively taxing) is not allowed in our programme and is actively monitored by the attendings and the supervising senior resident on shift. Prior research by our group has shown that there is a significant decrease in the number of patients seen per hour as the shift progresses, which is mirrored in our data14 15 (online supplementary figure 1). The total number of patients seen by shift hour, patients admitted by shift hour and proportion of patients admitted by shift hour is shown in online supplementary table 2.
Supplemental material
Supplemental material
We also modelled the following covariates: patient age and gender, the hour of day (modelled as a continuous integer), and the patient’s illness severity (represented by the Emergency Severity Index (ESI), a measure assigned at triage). Our ED has a predictable daily rhythm in terms of overall patient volume; we therefore used the hour of day as a proxy for ED arrival volume.
Statistical analysis
We first carried out bivariate analyses to assess the trends of association between our three classifications of shift hour and the likelihood of admission. To obtain an estimate of the effect of shift hour on the likelihood of admission, adjusted for potential confounders, we used a generalised estimating equation with Poisson error, a log link and exchangeable correlation structure, with physician as the clustering level. Using robust methods for estimating standard errors was meant to account for correlation within the physicians in the study, independent of time. An exchangeable correlation structure was used because we assumed observations within physicians were equally correlated. The analysis was intended to generate population-averaged estimates to inform department-level policies on physician staffing, adjusted for differences in individual provider patterns. To estimate the effect on different levels of illness severity, we also performed a secondary analysis of effect of shift hour on likelihood of admission in the last hour, stratified by ESI.
There are two layers of potential clustering in this data set: patient-level and physician-level. Regarding patient-level clustering, in our de-identified data set, each observation represents a unique ED visit, but not necessarily a unique patient; the same person may have visited the ED multiple times during the study period. Because of the de-identified data set, we could not account for this in the analysis. Regarding physician-level clustering: because individual physicians have routine practice patterns, the decision to admit by one person is not independent between observations. To account for this, we used a generalised estimating equation as above.
A value of p≤0.05 was used as the threshold of statistical significance. Our three classifications of shift hour were analysed independently in three different models each with age, sex, ESI (a common estimate of illness acuity, made in triage; scores range from 1 to 5, with 1 being most severe) and 24 hours clock time. Coefficients were exponentiated to estimate relative risks for each covariate and their 95% CIs.16 Patients with missing data for any of the covariates included in the regression models were excluded from analysis (n=191). We also calculated the attributable risk of admission (eg, the attribution fraction—the proportion of all cases that are the result of an exposure of interest).16 The purpose of this analysis was to estimate the number of admissions that could be attributed to the shift hour effect, if present. All analysis was conducted using R (R Project for Statistical Computing, www.r-project.org).
For sample size, as no prior literature in this area exists, the estimated effect size (and thus the sample size needed) was difficult to estimate. Six years of data contained sufficient information to measure all exposures, outcomes and covariates of interest.
Results
There were 294 031 ED visits during the time period, from which 191 observations were excluded due to missing data; of the included patients, 113 436 (38%, 95% CI 38.4% to 38.8%) were admitted to the hospital. The main features of the ED under study and the representative populations are shown in table 1. From the 294 031 ED visits, 40% were admitted in hour 8 (5098/12749, 95% CI 39.1% to 40.8%) and 40.3% were admitted in hour 9 (1186/2940, 95% CI 38.6% to 42.1%) (online supplementary table 2); shifts during the time period studied were either 8 hours or 9 hours in length. After taking into account clustering at the physician level and adjusting for age, ESI, sex and 24 hours clock time, we found a small effect of patients seen in the last hour of a clinician’s shift being around 3% more likely to be admitted compared with any other time in the shift (RR 1.03, 1.01–1.06). Those seen in the last quarter of the provider’s shift were approximately 2% more likely to be admitted when compared with patients seen in the first quarter of the shift (RR 1.02, 1.01–1.03). We also found that patients seen in later shift hours were more likely to be admitted, relative to their likelihood in the first hour; the relative risk of admission (compared with hour 1) is 1.04 in hour 6, (1.02–1.05), 1.03 in hour 7 (1.01–1.05), 1.04 in hour 8 (1.01–1.06) and 1.06 in hour 9 (1.013–1.101). Detailed results of this analysis can be seen in table 2 and figure 1. Based on our estimate of the total attributable risk, we estimate that 78 additional hospitalisations occurred due to being seen in the last hour of a clinician’s shift during the time frame studied.
Relative risk of admission by shift hour. Hour 1 is the reference category. Relative risk here is estimated from the generalised estimating equation (GEE) and adjusted for age, gender, 24 hours clock and ESI. Shaded region shows the 95% confidence limits. ED, emergency department; ESI, Emergency Severity Index.
Baseline characteristics of the patients, emergency department, providers and hospital examined in this study
Results of generalised estimating equation showing the primary exposure (main predictors) of last hour of shift (relative to the prior n hours), last quarter of shift (relative to the first three quarters) and shift hour (relative to the first hour)
In all models, several other results were constant. Female gender was significantly associated with lower likelihood of admission in all models, but the effect size was minimal (RR 0.92 in all models). The 24 hours time was associated with increased risk of admission in all models, but the effect size was minimal (RR 1.003 in all models). Finally, lower ESI (where lower numbers indicate more acute illness) was associated with increased risk for admission in all models, as would be expected.
In the ESI-stratified model, we found that the estimated effect of being seen in the last hour of a clinician’s shift was different across ESI categories. These results are summarised in online supplementary table 1. With respect to last hour of the shift, patients with an ESI of 2–3 saw the only statistically significant effect (RR 1.07, 1.03–1.10); shift hour was not significant in the ESI 1 and ESI 4–5 models. With respect to the last quarter of the shift, there was a statistically significant effect in patients with ESI 1 (RR 1.03, 1.01–1.04) and ESI 2–3 (RR 1.03, 1.02–1.05). Finally, with respect to shift hour (relative to hour 1), there was a statistically significant increased likelihood in hours 6–9 for ESI 1, hours 3–9 for ESI 2–3, and hours 7 and 9 for ESI 4–5.
Supplemental material
Discussion
In this study of 294 031 ED visits, we found a significant association between the hour of a provider’s shift during which a patient is electronically claimed and the likelihood of that patient being admitted. The likelihood of being admitted increases in the last hour, last quarter and later hours of a provider’s shift, reaching a peak relative risk of 1.06—corresponding to an approximately 6% increased likelihood of a patient being admitted by a provider who sees that patient in the ninth hour of a shift. This effect remains when the analysis is stratified by illness severity, as seen in online supplementary table 1. However, the last quarter of the shift accounts for less than 25% of a provider’s patient volume, on average, so the total number of increased admissions attributable to the shift hour effect is relatively small. Though other studies have examined provider-level variation in decision-making, to our knowledge, our study is the first to examine temporal changes in admitting behaviour in ED providers.
There are multiple potential contributing factors, which can be divided into provider-level and systems-level factors. With respect to provider-level factors, prior studies have looked at provider-level variation in admissions. Such variation has been studied with respect to all admissions within a health system,3 and several specific conditions, including chest pain,4 pneumonia5 and skin infections.6 Our study asks a somewhat different question: Is the same provider more likely to admit a patient later in the shift than she would have earlier on in the shift? In prior studies of variation, authors hypothesise that differences in provider education, training, risk tolerance, use of decision rules and fear of malpractice likely contribute to the variation. Some of those factors are immutable (eg, an attending’s residency experiences and education); however, even considering the mutable factors (fear of malpractice, risk tolerance and so on), those features are highly unlikely to change as a shift progresses. To understand the effect seen in our study, we must ask not about what features explain between-provider variation, but rather what factors explain within-provider variation.
The authors believe that the most likely explanation of our findings is the cumulative effects of fatigue and decreased cognitive reserve experienced by providers as the shift progresses. Conceptually, discharging a patient likely requires more cognitive resources than admitting a patient; the provider must review the data again to ensure discharge is safe and no further testing or treatment is needed, counsel the patient, compose discharge instructions, write prescriptions and arrange follow-up. If the patient is admitted, the patient’s care is transitioned to another physician with no additional input required. Admission also adds a layer of safety, as at least one additional physician will review the case, and admission guarantees a longer period of observation.
Systems-level factors may also play a role in providers’ decision to admit later in the shift. For example, hospital size7 8 as well as ED crowding9 10 have been shown to increase rates of admission. ED crowding might partly explain our findings, though our data do not reflect the degree of crowding (see limitations below). Additionally, department-level choices around staffing may also play a role, especially if mismatched to patient volumes. For example, if night-time volumes are high, but the department transitions from two to three providers to a single provider, the effects of ED volume on provider decision-making might become more pronounced. Even if a systems-level factor does partly explain our findings, the authors still believe that the ultimate impact of systems-level factors (for example, ED crowding or administrative pressure to reduce wait times) happens at the level of the individual provider deciding to admit the patient.
The more important finding is that, though the relative risk increases and is statistically significant, the absolute number of patients affected is small (we estimate an additional 78 additional admissions in the last hour, for the time period studied, relative to 293 840 ED visits and 115 199 admissions). A prior single-centre retrospective study by Chan suggests that ED physicians increase the number of admissions relative to discharges and are more likely to use additional resources at the end of the shift—in his interpretation, for purposes of increasing ‘leisure’ and ‘slacking off’ to avoid uncompensated work at the end of the shift.11 Though the measured effect in our study is statistically significant, it is not clinically significant given the small number of 78 additional admissions (translating to roughly 1 fewer admission per month). At least with respect to admissions, our results do not support immediate changes in staffing patterns. Further multicentre research is needed to clarify the overall impact of shift hour on admitting patterns.
Regardless of the impact on staffing, our findings still have relevance, as they may offer some insight into the cognitive effects of being on-shift, as described above. Admission or not is a relatively crude outcome, similar to mortality, and our results suggest that both: the effects of later shift hour on provider performance and other patient-level outcomes potentially affected by fatigue later in the shift require further study. Our operations group has shown that the volume of new patients seen by both residents and attending physicians decreases as shift hour increases.14 15 This is presumably because, as the shift progresses, providers accumulate active patients and tasks related to their care, limiting their ability to begin caring for new patients. But what if fatigue-related cognitive performance limits our ability to perform new work? Prior psychology research shows that mental fatigue is associated with declines in performance across multiple domains of executive function.17 Both multitasking and forced task switching, which typically increase as shift duration (and patient burden) increases, have been shown to decrease performance.18 When viewed in light of the findings of our study, these lines of evidence suggest that emergency providers may develop cognitive fatigue that could affect decisions about resource utilisation.
The results of this investigation must be interpreted within the limitations of our data set and methodology. First, this was a single-centre study; it should be replicated at other centres before being generalised. Second, it was performed using data from a large tertiary care centre; there may be different results if this study were performed using data from a community setting, though such a study would also need to take into account interhospital transfer as a potential outlet for cognitive easing. Third, we adjusted for ESI but did not have data available regarding the patients’ underlying comorbidities (such as that required to calculate the Charlson Comorbidity Index). This is an important consideration; the effect of shift hour on admission might be less pronounced in patients with more underlying comorbidities, and this should be investigated in a future study. However, comorbid illness would not explain a temporal pattern in physician admitting patterns; in fact, patients with multiple comorbidities would be less likely to experience temporal variation in admitting patterns, because their underlying complexity would make them more likely to be admitted regardless of timing of presentation. Fourth, we did not specifically account for the effect of chief complaint; similar to the issue with comorbidities, we expect that certain subgroups might be differentially affected, but that specific chief complaints would not explain temporal variation in provider behaviour. Fifth, we accounted for provider behaviour using clustering (as described in the Methods section), but we assumed all observations in the data set (eg, patient visits) were independent. However, some visits are likely repeat visits by the same patient, in which case those specific observations would not be independent. Given the large number of observations, this degree of correlation might affect the variance of our estimates, but is less likely to affect the magnitude or statistical significance. Sixth, we were unable to include total ED volume (eg, crowding) in our model as a time-varying variable. Given the association between ED crowding and admissions decisions shown in other studies, it will be important to include this explicitly in future work. Seventh, we included provider shift hour and 24 hours time, but did not include provider shift type or the number of providers on shift (which might be more generalisable, given the institution-dependence of shift type). Eighth, we did not include the total number of primary patients or sign-out patients being cared for by providers; we suspect these may play a role, as caring for more sign-out patients has been shown to decrease the amount of new work done by a provider. Ninth, we used the attending as the provider of interest in this study, but the resident likely plays a role as well; we intend to study this in future work. Finally, any cohort study risks selection bias; we attempted to ameliorate this by including all patient visits from the period in question.
Our study has several potential implications. First, it raises the question of what features may be contributing to this change in decision-making. The authors’ experience as practising emergency physicians suggests that fatigue may be an important factor. If our findings are independently verified, the underlying factors should be explicitly and independently studied. Second, if these findings are validated in multicentre studies, it suggests that we may need to tailor provider shifts to account for changes in provider decision-making. At some centres, shifts are designed and staggered such that providers see patients in stages of decreasing acuity as shifts progress. This staffing structure has the advantage of placing less cognitively taxing cases at the end of the shift. It also has the interesting helpful effect of decreasing sign-outs, because the time to disposition decreases for lower-acuity cases.19 20 Prior research has shown that an increased sign-out burden decreases the number of new patients seen and total number of revenue value units generated per shift.21
Multiple areas of future research are needed to further investigate this effect. First, future research should study these findings in multiple centres, as well as investigate the effects of shift hour on admissions (and transfers) in the community setting relative to the tertiary care academic setting. Second, studies of temporal variation should examine the impact of residents on patient care; for example, one could include postgraduate year (categorical variable), or time in training (postgraduate year represented as a continuous variable), as a factor in models. Third, future studies should also incorporate systems-level features into the model. Examples of these features include ED crowding, as well as the number of additional providers (resident and attending) who are simultaneously on shift. Given the complexity of accounting for shift type, we suspect that it will be more straightforward and generalisable to study the number of simultaneous providers rather than shift type as the variable of interest. The effect of provider burden (eg, total number of primary and signed-out patients) on likelihood of admission should also be studied. Fourth, it would also be interesting to study the effects of shift hour when stratified for the 10 most common diagnoses for individual EDs. Fifth, there may be in impact of shift hour on decisions other than admission; for example, shift hour may (or may not) change the likelihood of a patient having additional laboratory testing or imaging studies ordered. We intend to evaluate many of these lines of inquiry in future work. Finally, if the research above reveals temporal variation to be an important factor in patient outcomes, further work will need to examine the causes of the variation.
Conclusions
In a single academic centre, our study showed that the provider shift hour during which a patient is seen increases the likelihood of that patient being admitted by as much as 6% as shift hour increases. We believe this is due to the cognitive effects of progressive time on shift. Several limitations should be considered when interpreting these findings, as enumerated above. Future work should validate these findings in multiple centres, explore whether similar effects are present in community sites and explicitly investigate the cognitive effects suggested by our results.
Footnotes
Funding This study was funded by Massachusetts College of Emergency Physicians (Resident Research Grant).
Competing interests None declared.
Patient consent for publication Not required.
Ethics approval This study was deemed IRB exempt by the IRB at Beth Israel Deaconess Medical Center, Protocol #: 2017P000593, as de-identified data were used.
Provenance and peer review Not commissioned; externally peer reviewed.
Data availability statement Data are available upon reasonable request.