AI model predicts patient outcomes from surgical gestures and provides insights into explainability

Abstract

Effective surgical training requires relatively immediate feedback as to outcomes. This makes surgical learning problematic as some surgical outcomes may take months or years to become apparent. The sequence of surgical gestures, the smallest discrete actions of surgery, during the nerve-sparing step of robot-assisted radical prostatectomy has been used to predict 1-year erectile function (EF) outcomes after surgery. To improve this prediction and extract clinically meaningful insights, we describe the addition of anatomic and functional context to surgical gestures. We analyzed surgical video of 147 patients at 5 surgical centers undergoing robotic-assisted radical prostatectomy. The addition of anatomic and functional characterization to surgical gestures improved model prediction of post-operative EF from 0.78 [95%CI: 0.60, 0.92] to 0.85 [95%CI: 0.66, 0.96]. Aggregated attention weight analysis identified novel gesture, anatomy, and function combinations contributing most to EF outcomes. The identification of these critical gestures provides a starting point for more data-driven training in clinical practice.

Ähnliche Arbeiten