Empirically Transformed Energy Patterns: A novel approach for capturing fNIRS signal dynamics in pain assessment.

The accurate assessment of pain in clinical settings is challenging due to its subjective nature. In this study, we used functional near-infrared spectroscopy (fNIRS) to measure brain activity by detecting changes in blood oxygenation. Leveraging the AI4Pain Grand Challenge dataset, we aimed to classify pain levels into No Pain (NP), Low Pain (LP), and High Pain (HP) categories using both binary (NP vs. HP) and multiclass (NP vs. LP vs. HP) approaches. This involved collecting a comprehensive dataset of fNIRS data from 65 subjects, with recordings from 24 channels. We proposed novel Empirically Transformed Energy Patterns to extract meaningful bio-information related to pain conditions. An optimised Ensemble Classifier, evaluated using leave-one-subject-out cross-validation, was employed to classify pain levels. Our results demonstrated that this approach outperformed traditional classifiers, achieving 91.41% accuracy in the binary task and 68.20% accuracy in the multiclass task, with high sensitivity and specificity. This study highlights the effectiveness of using optimised machine learning models with fNIRS data for precise pain level classification, which holds significant potential for improving pain management in clinical settings.