Machine Learning-Based Pain Intensity Assessment Using Multimodal Physiological Signals and Resampling Techniques in Clinical Datasets

This study proposes a multi-modal physiological signal-based machine learning model for three-level pain classification using the PainMonit Clinical Dataset collected in real clinical settings. Six physiological signals (BVP, EDA, EMG, Respiration) were analyzed, with 228 handcrafted features reduced to 43 via Recursive Feature Elimination. Five machine learning algorithms (XGBoost, AdaBoost, KNN, MLP, RF) were evaluated in combination with four resampling strategies (Raw, B-SMT, ENN, Hybrid), totaling 20 models. Performance was assessed using AUPRC, Recall, F1-score, and G-mean. Results showed RF, XGBoost, and KNN outperformed others, with RF-B-SMT achieving the best balance of recall and accuracy. SHAP analysis revealed EMG features as dominant predictors, followed by EDA and respiration metrics. This work addresses limitations of prior heat-stimulus-based studies by employing clinically collected multi-modal data and exploring advanced resampling methods to improve model robustness and clinical applicability.