Classifying clinical work settings using EHR audit logs: a machine learning approach

• Published in: The American journal of managed care Vol. 29; no. 1; pp. e24 - e30
• Main Authors: Kim, Seunghwan, Lou, Sunny S, Baratta, Laura R, Kannampallil, Thomas
• Format: Journal Article
• Language: English
• Published: United States Intellisphere, LLC 01.01.2023; MultiMedia Healthcare Inc
• Subjects: Anesthesia; Anesthesiology; Audits; Classification; Clinical medicine; Computational linguistics; Electronic Health Records; Electronic records; Emergency medical care; Health Personnel; Humans; Intensive care; Intensive Care Units; Language processing; Machine Learning; Medical colleges; Medical personnel; Medical records; Natural language interfaces; Patients; Physicians; Primary care; Workloads; Workweeks; Missouri
• ISSN: 1088-0224; 1936-2692
• DOI: 10.37765/ajmc.2023.89310

We used electronic health record (EHR)-based raw audit logs to classify the work settings of anesthesiology physicians providing care in both surgical intensive care units (ICUs) and operating rooms. Observational study. Attending anesthesiologists who worked at least 1 shift in 1 of 4 surgical ICUs in calendar year 2019 were included. Time-stamped EHR-based audit log events for each week were used to create event frequencies and represented as a term frequency-inverse document frequency matrix. Primary classification outcome of interest was a physician's clinical work setting. Performance of multiple supervised machine learning classifiers were evaluated. A total of 24 attending physicians were included; physicians performed a median (IQR) of 2545 (906-5071) EHR-based actions per week and worked a median (IQR) of 5 (3-7) weeks in a surgical ICU. A random forest classifier yielded the best discriminative performance (mean [SD] area under receiver operating characteristic curve, 0.88 [0.05]; mean [SD] area under precision-recall curve, 0.72 [0.13]). Model explanations illustrated that clinical activities related to signing of clinical notes, printing handoff data, and updating diagnosis information were associated with the positive prediction of working in a surgical ICU setting. A random forest classifier using a frequency-based feature engineering approach successfully predicted work settings of physicians with multiple clinical responsibilities with high accuracy. These findings highlight opportunities for using audit logs for automated assessment of clinician activities and their work settings, thereby affording the ability to accurately assess context-specific work characteristics (eg, workload).