Physiological Data Analysis Framework for Pain Prediction in Physical Rehabilitation

Predicting pain in physical rehabilitation is challenging due to subjectivity, patient variability, and self-report bias, especially in telerehabilitation. This study aims to determine whether machine-learning models based on heart rate (HR), heart rate variability (HRV), and peripheral oxygen saturation (SpO₂) can reliably detect clinically meaningful pain during real rehabilitation sessions, including home-based settings where self-report is least reliable; we hypothesized that these low-cost, non-invasive markers carry sufficient information to flag low-to-moderate pain episodes without relying on self-report. We combined these markers with machine-learning models. These markers were selected for their association with autonomic pain responses and ease of measurement with only two low-cost, non-invasive sensors (a wearable band providing HR and HRV, and a fingertip oximeter providing SpO₂) suitable for clinical and home-based rehabilitation. We evaluated linear regression (LR), random forest (RF), and artificial neural networks (ANNs) using data from 25 participants (aged 20-50) undergoing lower-limb rehabilitation. Signals acquired at 1 Hz were processed via temporal filtering, quality screening, and three missing-value strategies (interpolation, zero imputation, deletion) before normalization and training. LR showed limited predictive power. RF achieved 97.77% accuracy in detecting low-pain episodes, and balanced per-class performance under deletion (76.64%). ANN models contributed a more balanced three-class profile on interpolated data but remained sensitive to class imbalance. Given high-pain scarcity in supervised therapy and underreporting at home, reliable detection of low-to-moderate pain enables timely therapy adjustments. Unlike prior studies using experimentally induced pain, this work captured naturally occurring pain during real rehabilitation, making findings applicable to clinical and telerehabilitation contexts. Physiology-based models with low-cost sensors show promise for personalized rehabilitation, improving adherence and enabling proactive adjustments without added complexity.