Using Time Series Clustering to Segment and Infer Emergency Department Nursing Shifts from Electronic Health Record Log Files

Abstract

Few computational approaches exist for abstracting electronic health record (EHR) log files into clinically meaningful phenomena like clinician shifts. Because shifts are a fundamental unit of work recognized in clinical settings, shifts may serve as a primary unit of analysis in the study of documentation burden. We conducted a proof- of-concept study to investigate the feasibility of a novel approach using time series clustering to segment and infer clinician shifts from EHR log files. From 33,535,585 events captured between April-June 2021, we computationally identified 43,911 potential shifts among 2,285 (74.2%) emergency department nurses. On average, computationally-identified shifts were 10.6±3.1 hours long. Based on data distributions, we classified these shifts based on type: day, evening, night; and length: 12-hour, 8-hour, other. We validated our method through manual chart review of computationally-identified 12-hour shifts achieving 92.0% accuracy. Preliminary results suggest unsupervised clustering methods may be a reasonable approach for rapidly identifying clinician shifts.

Figure 1.

Schematic of our two-step time series clustering approach to automatically identify shifts using raw EHR log files. In this scenario, EHR event timestamps for two nurses, depicted as spheres (teal and red respectively), are subsetted at the individual-nurse level and clustered into shifts based on timestamp density using DBSCAN (gray boxes) in Step 1. In Step 2, rule-based logic conditioned on the first timestamp identified in the time series is applied to classify clusters into day, evening and night shifts.

Figure 2.

Distribution of EHR events that nurses logged per day, April 2021 (M=median; Q=quartile; Min=minimum; Max=maximum)

Figure 3.

Distribution of time intervals between EHR events logged among nurses in hours, April 2021*

Figure 4a.

Distribution of shift duration among computationally-identified nurse shifts*

Figure 4b.

Distribution of shift start times among computationally-identified nurse shifts

Figure 5.

Process model of our two-step approach starting with raw EHR log file data extraction to the validation of computationally-identified shifts using Kronos data.