A Multi-Level Physiological Validation Framework for HRV-Derived Surrogate Stress Labels in Surgical Environments

Continuous monitoring of surgeons' stress has the potential to improve surgical performance, training, and patient safety. However, developing reliable machine learning models for stress estimation is challenging because continuous psychological ground truth is impractical to obtain during surgery. Surrogate-label learning has emerged as a promising alternative, yet the physiological validity of the generated labels and the resulting model predictions remains largely unexplored. This study presents a multi-level physiological validation framework for evaluating personalized heart rate variability (HRV)-derived surrogate stress labels and corresponding deep learning models for continuous intraoperative stress estimation. The framework comprises two complementary validation levels. The first evaluates the physiological consistency of personalized clustering-derived surrogate labels against HRV-derived stress profiles using representative surgical cases, surgical workflow analysis, and correlation analysis. The second assesses the agreement between surrogate labels and convolutional neural network (CNN) and long short-term memory (LSTM) model predictions using prediction error metrics and Bland–Altman analysis across multiple temporal sequence lengths. Validation was performed using wearable electrocardiogram (ECG) recordings collected from 27 colorectal surgical procedures performed by seven surgeons. The surrogate labels demonstrated strong agreement with the HRV-derived reference (Spearman's ρ = 0.841, p = 0.036), while both CNN and LSTM models accurately preserved the physiological stress patterns, exhibiting moderate-to-strong correlations, low prediction errors, and acceptable agreement across all evaluated sequence lengths. The proposed framework extends conventional model evaluation by emphasizing physiological consistency in addition to predictive performance and provides a systematic methodology for validating surrogate-label-based physiological monitoring systems when continuous psychological ground truth is unavailable.