Predicting task performance in robot-assisted surgery using physiological stress and subjective workload: a case study with interpretable machine learning

Abstract

Robot-assisted surgery (RAS) enhances surgical precision and extends surgeons' capabilities. However, its effects on the cognitive and physical states of surgeons remain poorly understood. It is essential to investigate the workload and physiological stress surgeons experience during RAS. This case study employs a neuroergonomic approach to explore how these factors relate to task performance. A single expert surgeon performed simulated surgical tasks under systematically varied conditions (noise level, surgical posture and task type) to elicit variations in stress and workload. During the tasks, multiple physiological signals were recorded, including electroencephalography (EEG), electromyography (EMG), heart rate (HR), and electrodermal activity (EDA). Subjective workload was also assessed using the NASA-TLX and SURG-TLX. Several classification models, including CatBoost, random forest, logistic regression, and support vector machines, were trained to predict task performance. Among them, CatBoost demonstrated the highest predictive accuracy (79.5%) and achieved an area under the curve (AUC) of 0.807. The model interpretation was conducted using SHapley Additive exPlanations (SHAP). The analysis revealed that subjective workload, mean HR, and muscle activation were the most influential predictors. EEG-related features contributed variably across conditions. This study shows that integrating subjective assessments with physiological measures can effectively predict surgical task performance under stress.