HRV-Based Surrogate Classification in Surgical Environments Using Personalised Labelling and Sequence Modelling

Accurate monitoring of stress and workload in surgical environments is essential for improving performance and patient safety. Physiological signals, particularly heart rate variability (HRV), offer a promising approach for objective and continuous stress assessment. However, existing studies often rely on population-level models, limited temporal analysis, and evaluation strategies that may not reflect real-world generalisation. This study proposes an end-to-end pipeline for HRV-based classification of HRV-derived surrogate stress-like states in surgical settings, combining personalised feature extraction, surrogate label generation, and sequence-based deep learning. HRV data were collected from multiple surgeons during real colorectal procedures using wearable sensors and processed into minute-level feature representations. Stress labels were generated using subject-specific thresholds to account for individual physiological differences. Sequence-based models, including convolutional neural networks (CNNs) and long short-term memory (LSTM) networks, were evaluated across multiple temporal window lengths and validation strategies. The results show that mean-based Standard Deviation of Normal-to-Normal intervals (SDNN) provided the most consistent separation of physiological states. The highest-performing models were most frequently associated with temporal windows of 10–20 min, while CNN and LSTM architectures achieved comparable performance. Notably, in this dataset, leave-one-subject-out (LOSO) validation produced performance that was comparable to, and in some configurations higher than, pooled random splits, highlighting the influence of evaluation strategy and the potential benefits of personalised labelling. Overall, this work suggests that personalised HRV-based modelling and temporal sequence analysis may provide a feasible approach for identifying HRV-derived stress-like physiological states in real surgical environments. The findings also emphasise the importance of rigorous evaluation design when developing physiological machine learning systems.