Background: Provider-level practice variation is common and may impact clinical outcomes, such as length of stay and imaging ordering. However, it is unclear which provider characteristics lead to variations in practice. Using individual performance data from a provider-level dashboard, we aimed to evaluate variability in clinical performance and predictors of high or low performance.
Methods: This was an observational study of all admissions to the hospital medicine service at an urban tertiary academic hospital for between July 1, 2020 and June 30, 2021. All hospitalists that worked ≥20 days during this time period were included in analysis. Hospitalists were divided into four performance quartiles based on each of three selected quality metrics: 1) individualized length of stay (LOS), 2) appropriate glucose control (measured as the percent of days patients were above or below pre-set thresholds of appropriate blood glucose levels), and 3) diagnostic spending (dollars spent on imaging and lab orders per patient per day). Individual physician characteristics were collected from publicly available information. We compared physician characteristics of those who were in the highest performing quartile (quartile 1, q1) to the lowest performing quartile (quartile 4, q4) using chi-squared and t-tests for each metric. We ran multivariate logistic regression adjusting for case-mix index (CMI), an indicator of relative case complexity. Hospitalists at this medical center spend variable amounts of time working in any of the eight teaching teams and/or ten direct-care (non-teaching) teams. Physician time spent on the direct-care hospitalist service was analyzed as a binary variable as either ≥50% or < 50% of total days worked spent on the direct-care service. Dashboard engagement was measured as the mean number of times logged in to the performance dashboard website per month.
Results: A total of 95 hospitalists met inclusion criteria for the study. Outcome measures for providers in q1 (highest performing) and q4 (lowest performing) are summarized in Table 1. Multivariate logistic regression results, adjusted for CMI, are shown in Table 2. Between LOS quartiles 1 and 4, increased time since completing residency was a borderline significant (p=.086) predictor of highest performance (shorter length of stay), with each year post-residency corresponding to a 10% lower odds of being in q4 as opposed to q1. For glucose control, dashboard engagement positively correlated with improved glucose control, with each login associated with a 35% lower odds of appearing in q4 over q1 with borderline significance (p=0.076). For diagnostic spending, spending ≥50% days worked on the direct-care service was correlated with a statistically significant (p=0.037) 95% lower odds of appearing in q4 over q1.
Conclusions: We found statistically significant, clinically meaningful differences between the highest and lowest performing quartiles of physicians in all three relevant quality measures. Time since residency, degree of engagement with the performance dashboard, and exposure to direct-service can correlate with greater odds of appearing in the top performing quartile depending on the measure. These findings highlight possible quality improvement opportunities and can serve as an initial guide in targeting practice improvement.
