A Machine Learning Approach for Pain Detection Using Physiological Signals

Abstract

This study aims to develop a machine-learning model for classifying the presence of pain using 5.5-second recordings of electrocardiogram (ECG), galvanic skin response (GSR), and electromyogram (EMG) signals. Two approaches were used: a single model and a multi-model (ensemble learning). The ensemble approach utilized four aggregation methods: majority voting, soft voting (stationary threshold and adaptive threshold), and weighted voting. Various machine learning algorithms, including SVM, XGBoost, KNN, Logistic Regression, and Random Forest, were compared regarding accuracy, precision, recall, F1 score, and computational efficiency. The findings highlight the strengths and weaknesses of each algorithm and propose the optimal model for real-time pain detection. The results revealed highly promising accuracy with the multi-model approach using adaptive threshold voting, achieving the highest performance at 86.84 %. The multi-model approach with stationary threshold voting was closely followed, attaining an accuracy of 86.26%. The XGBoost single model achieved an accuracy of 85.98%. Among the PCA models, the soft adaptive threshold voting performed best at 85.40%, followed by the soft static threshold at 83.86%.