Optical Flow-Derived Motion Biomarkers and Explainable Artificial Intelligence for Quantitative Assessment of Robotic Surgical Expertise.

Abstract

BackgroundObjective surgical skill evaluation remains challenging in robotic-assisted surgery. Existing artificial intelligence frameworks often depend on robotic telemetry or "black-box" deep learning models that reduce interpretability and translational accessibility.MethodsThe present study developed an interpretable computer vision framework for quantitative analysis of robotic surgical motion using dense optical flow-derived biomarkers extracted directly from surgical video recordings. Robotic suturing videos from the publicly available JIGSAWS dataset were analyzed using a Python/OpenCV-based pipeline. Quantitative motion biomarkers, including motion magnitude, temporal variability, jerk-related parameters, and smoothness indices, were extracted and evaluated across expert, intermediate, and novice surgical groups. The paired left and right stereo views were averaged to produce one feature vector for each independent trial. Statistical comparisons were performed using the Kruskal-Wallis test. Skill classification was evaluated using a class-weighted Random Forest classifier with Leave-One-User-Out cross-validation. Model predictions were interpreted using SHapley Additive exPlanations (SHAP).ResultsOptical flow analysis successfully extracted reproducible motion biomarkers associated with surgical expertise. Significant differences were observed in mean motion magnitude (P = .0035), mean absolute jerk (P = .00074), and the unnormalized smoothness index (P = .00074). Motion variability did not differ significantly among the groups after trial-level aggregation (P = .0769). SPARC showed the expected ordering of normalized smoothness, with the highest values among experts, but the group difference was not statistically significant (P = .0647). Principal component analysis demonstrated partial clustering of surgical expertise groups based on extracted motion dynamics. Under Leave-One-User-Out cross-validation, the Random Forest classifier achieved an overall accuracy of 38.5%, a balanced accuracy of 29.5%, a macro-average AUC of 0.401, and a weighted-average AUC of 0.430. Genuine out-of-fold SHAP analysis identified the smoothness index, mean absolute jerk, and mean absolute acceleration as the largest contributors to the model outputs.ConclusionsOptical flow-derived motion biomarkers provide an interpretable characterization of robotic surgical movement directly from video recordings without requiring robotic telemetry. However, the modest performance under user-independent validation indicates that these biomarkers should not yet be used as a standalone measure of surgical expertise. The proposed framework provides a transparent basis for developing and externally validating video-based surgical skill assessment methods.