Learning Manipulation Features for Quantitative Assessment and Skill-Level Classification in Robot-Assisted...

Robot-assisted vascular interventions demand precise manipulation within tortuous millimeter-scale vessels, where surgical outcomes remain critically dependent on surgeon skill. Conventional metrics such as completion time or navigation success are too coarse to capture the subtle, yet decisive control differences between experts and novices, while sensor-heavy external approaches obscure the surgeon's intrinsic dynamics. This article introduces a task-aware spectral modeling (TASM) framework that directly extracts skill signatures from joystick control signals, revealing how surgical expertise manifests in the time-frequency domain. In vivo rabbit experiments demonstrate that expert manipulation emerges as smooth, deliberate low-frequency rhythmic control, whereas novice performance degenerates into fragmented, high-frequency corrections. Spectral analysis indicates that approximately 96% of control energy lies below 5Hz, suggesting that low-frequency components may serve as an indicative spectral characteristic associated with expert-novice differentiation under similar experimental and task conditions. By deriving interpretable spectral descriptors, the framework achieves an average ROC-AUC of 0.992 across nine classifiers. Clinically, the proposed framework may support quantitative assessment of surgical skill and control stability in robot-assisted endovascular procedures, with potential value for identifying inconsistent control behaviors and informing future training and evaluation pipelines.