Measuring provider-level differences in perioperative workflow using computer vision-based artificial intelligence

Comprehensive Summary

This study aimed to evaluate the ability of a computer-vision based artificial intelligence (AI) system to properly and accurately segment events in the operating room (OR) in real time. For this study, total knee arthroplasty cases from September 2022 to March 2025 were used to collect data. A YOLO-based model, which splits computer vision into a grid to predict possibilities in each individual square, was used to identify patients, staff, and equipment. The AI model then segmented the cases into the following main events by identifying which staff were actively working: anesthesia induction, patient preparation, final preparation, active procedure, post operation, patient exit, room cleanup, and room setup. The computer vision was evaluated for F1 score, accuracy, precision, recall and median time error against manual annotations. The AI model boasted impressive performance across all evaluations, with F1 scores, accuracy, precision and recall values all ≥ 0.97, and median time errors between 6.4 and 16.9 seconds. This indicates that the model was able to accurately and consistently identify key events as they happened during a surgical procedure.

Outcomes and Implications

Operating room time is a significant expense for hospitals, with every minute being costly. Therefore, it is a priority for hospitals to cut down on operating room times wherever possible without compromising patient care. Dividing case durations into segments has been shown to decrease operating room inefficiencies, however it is challenging for operating room nurses to take these additional notes during a case. This AI model allows for hospital staff to evaluate OR efficiency quickly and easily, being able to tell exactly where cases run longer than they should. However, one limitation of this study is that it did not consider patient variability, such as age and body type. Some variation in the results may be attributed to these factors, and future studies should evaluate these variables in their methodology.