Strategies for Assessing Pain Using Multimodal Pain Detection

Pain assessment is crucial for effective medical care, particularly in cases where patients are unable to communicate their discomfort, such as infants or individuals with cognitive impairments, yet it remains challenging due to its subjective nature. Understanding and effectively managing pain are essential for improving the quality of life for individuals experiencing both acute and chronic pain. This paper presents a comparative study of various machine learning and deep learning methods for pain assessment with different feature selection techniques, utilizing the BioVid heat pain dataset with a focus on Electrocardiogram (ECG) and Skin Conductance Level (SCL) signals. The given dataset consists of around 90 participants subjected to heat stimuli with 4 different levels of heat and records their response with physiological signals. A feature extraction process was applied on each signal then deployed for evaluation on machine learning model: Random Forest classifier and SVM, also on deep learning models: LSTM and CNN-LSTM. Each model was trained and tested across the four distinct levels (PA1, PA2, PA3, PA4) and baseline (BL), achieving accuracies of 75.39% for Random Forest, 80.75% for SVM, 76.99% for LSTM, 75.54% for CNN-LSTM. Additionally, the models were evaluated to classify all 5 levels of pain together. In response to enhance the performance, two feature selection methods were applied, Boruta and one-way ANOVA, and compared to a non-features selection appliance.