Identifying Electronic Health Record Tasks and Activity Using Computer Vision.

Abstract

Background: Time spent in the electronic health record (EHR) is an important measure of clinical activity. Vendor-derived EHR use metrics may not correspond to actual EHR experience. Raw EHR audit logs enable customized EHR use metrics, but translating discrete timestamps to time intervals is challenging. There are insufficient data available to quantify inactivity between audit log timestamps.

Methods: We propose a computer vision-based model that can 1) classify EHR tasks being performed, and identify when task changes occur, and 2) quantify active-use time using session screen recordings of EHR use. We generated 111 minutes of simulated workflow in an Epic sandbox environment for development and training and collected 86 minutes of real-world clinician session recordings for validation. The model used YOLOv8, Tesseract OCR, and a predefined dictionary to perform task classification and task change detection. We developed a frame comparison algorithm to delineate activity from inactivity and thus quantify active time. We compared the model's output of task classification, task change identification, and active time quantification against clinician annotations. We then performed a post-hoc sensitivity analysis to identify the model's accuracy when using optimal parameters.

Results: Our model classified time spent in various high-level tasks with 94% accuracy. It detected task changes with 90.6% sensitivity. Active-use quantification varied by task, with lower MAPE for tasks with clear visual changes (e.g., Results Review) and higher MAPE for tasks with subtle interactions (e.g., Note Entry). A post-hoc sensitivity analysis revealed improvement in active-use quantification with a lower threshold of inactivity than initially used.

Conclusion: A computer vision approach to identifying tasks performed and measuring time spent in the EHR is feasible. Future work should refine task-specific thresholds and validate across diverse settings. This approach enables defining optimal context-sensitive thresholds for quantifying clinically relevant active EHR time using raw audit log data.