Interpretable Surgical Skill Assessment: A Feedback-Generative Framework Using Kinematic and Force Data Publisher



Rastaghi A ; Ahmadi M J ; Asadi Khameneh E ; Heidari R ; Riazi Esfahani H ; Tavakoli M ; Mohammadi S F ; Taghirad H D
Source: Results in Engineering Published:2026

Abstract

Surgical skill assessment is critical for patient safety and effective training, yet prevailing methods often lack objectivity and educational value. Expert based ratings are subjective and time consuming, while many automated approaches provide opaque scores with limited instructional relevance. This study proposes a feedback oriented framework that leverages synchronized kinematic and force measurements to derive clinically meaningful insights by integrating expert defined knowledge with temporal patterns learned directly from data. Manually defined metrics were interpreted using a Support Vector Machine, while data driven temporal patterns were simultaneously captured through a multivariate Shapelet Transform combined with an enriched Random Forest classifier. To maximize the benefits of both approaches, their outputs were integrated using an ensemble meta model based on XGBoost. The framework was evaluated in a proof of concept deep vitrectomy study using the ARASH:ASiST haptic system, with seven cadaveric eye recordings acquired under near clinical conditions, and its generalizability was further examined through a feasibility analysis conducted on the public JIGSAWS benchmark. Classification performance of each phase segment was better on the JIGSAWS benchmark than on the locally collected vitrectomy dataset, attributable to larger sample sizes, balanced class distributions, and simpler surgical tasks. Quantitative results were translated into structured narrative feedback using large language models, and the feedback was qualitatively reviewed by clinical experts for relevance and clarity. Expert review confirmed that the generated feedback aligns with clinically accepted interpretations of surgical performance and supports the framework’s feasibility, potential generalizability, and the need for larger scale validation and quantitative benchmarking in future studies. © 2026 The Author(s).