Validation of Surgical Pleth Index for Objective Pain Assessment Using Photoplethysmography and Electrodermal Activity

This is the authors' abstract. We don't add key points for this paper.

Objective pain assessment remains a major challenge, particularly for non-communicative patients. This study validates the Surgical Pleth Index (SPI) within an interpretable machine learning framework using synchronized photoplethysmography (PPG) and electrodermal activity (EDA) signals. We addressed two clinical tasks under controlled thermal stimulation: pain detection (no pain vs. pain) and pain intensity differentiation (low vs. moderate)—a more nuanced challenge than typical baseline vs. extreme pain contrasts. Five machine learning algorithms were evaluated using 20 extracted physiological features. SHapley Additive exPlanations (SHAP) guided feature reduction to establish parsimonious models. A LightGBM-based model demonstrated superior discriminative ability and robust generalization under strict leave-one-subject-out cross-validation (LOSO-CV). For pain detection, a 6-feature model yielded an average area under the ROC curve (AUC) of 0.9580, sensitivity of 0.8850, specificity of 0.9240, accuracy of 0.9120, and an F1 score of 0.8890. For intensity differentiation, a 9-feature model achieved an AUC of 0.8250 and specificity of 0.8350. Ablation and SHAP analyses confirmed that SPI contributes synergistically beyond its individual components, highlighting its relevance in conscious pain assessment. Compared to resource-intensive deep learning, this framework offers strong estimation, interpretability, and ultra-low inference latency (0.0023 μs per sample), making it highly suitable for real-time wearable monitoring systems.