1247 Expert Surgeon-Level OSATS Rating Using Machine Learning for Surgical Performance Assessment

INTRODUCTION: Real-time feedback can improve surgical skills, but current gold-standard approaches based on expert-provided Objective Structured Assessment of Technical Skills (OSATS) ratings are time-consuming, impractical, and unavailable to most surgeons and trainees. An artificial intelligence (AI) system can provide real-time, automated OSATS ratings which could bridge this gap. METHODS: Data from 532 tasks were recorded during medical students’ brain tumor surgery simulation training at the Neurosurgical Simulation and Artificial Intelligence Learning Centre, McGill University. These videos were rated by two blinded experts using OSATS. This data was split into three: training, validation, and testing datasets, for the development and independent verification in training the algorithm. A long short-term memory network was trained. RESULTS: Despite variations between expert ratings (cohen’s weighted kappa = .09), our machine learning algorithm achieved 78% accuracy overall and 79% accuracy in independent testing in differentiating between good and poor performance. The algorithm achieved 73% accuracy in predicting the OSATS ratings within a one-point range. CONCLUSIONS: Our work presents an automated real-time AI system that replicates expert decisions in surgical performance assessment. Provided with enough examples of good and poor surgical performance, machine learning systems may learn and emulate experts’ opinions. These systems may provide real-time expert-level feedback during procedures to inform the trainees about poor performance, help mitigate risks, and improve operative outcomes.