Individualized Pain Detection Models Using Physiological Signals from Wearable Devices

Chronic pain is inherently subjective, varying across individuals in perception and physiological response. To address this challenge, we develop individualized machine learning models for pain detection using physiological signals collected from a wrist-worn wearable device. These signals include electrodermal activity (EDA), photoplethysmography (PPG), heart rate variability (HRV), and skin temperature. Self-reported pain levels serve as ground truth labels for training our models. This study is part of an ongoing IRB-approved clinical trial, with data collected from more than 10 chronic pain participants to date, and additional participants being enrolled. Our approach involves extensive signal processing to extract relevant features, including time-domain, frequency-domain, and non-linear characteristics from each physiological modality. We employ machine learning techniques such as Easy Ensemble with XGBoost, which leverages feature-based learning for interpretable pain classification, as well as an encoder-decoder neural network to learn latent representations of pain states. Given the variability in pain perception, we train individualized models for each patient, allowing for more precise detection compared to population-level models. Our models demonstrate strong performance in pain classification, achieving high sensitivity (true positive rate) and specificity (true negative rate) across multiple participants. Preliminary results indicate average accuracy of 85% in distinguishing pain states, highlighting the effectiveness of personalized modeling. This work integrates AI-driven methods with wearable health monitoring, enabling real-time, personalized pain assessment, and contributes to the advancement of AI in medical devices by providing an objective, data-driven approach to chronic pain management