Connectivity-Based Pain Recognition from fNIRS: Parsimonious Subject-Independent Classification

Accurate pain assessment remains challenging, because pain is subjective, and common clinical tools rely on self-report or observation. Functional near-infrared spectroscopy (fNIRS) provides a non-invasive window into cortical haemodynamics and shows promise for objective pain assessment. In this study, we propose a subject-independent framework that models pain states using functional and effective connectivity features derived from fNIRS signals. Using the AI4PAIN dataset (65 participants, 24 channels), we extracted correlation, partial correlation, coherence, and Granger causality matrices from haemodynamic signals (HbO₂, HHb, and HbT), yielding a comprehensive connectivity feature space. A three-class classification task (No Pain vs. Low Pain vs. High Pain) was evaluated using leave-one-subject-out (LOSO) cross-validation. Within each training fold, we standardised features and applied mutual-information-based feature selection; a systematic sweep showed that a reduced feature set of 700 connectivity features achieves performance comparable to the full 1380 features and constitutes the minimum configuration at which High Pain instances begin to be detected (High Pain recall ≈ 50%), whereas smaller feature sets produce near-zero High Pain recall. The proposed framework achieved best accuracy of 69.6%, with HHb achieving the highest accuracy (69.6%), followed by HbT and HbO₂. Analysis of frequently selected features revealed a predominance of directed Granger-causality connections, indicating stable and interpretable connectivity patterns associated with pain processing. These findings demonstrate the feasibility of connectivity-based subject-independent pain assessment and highlight the potential of fNIRS for objective clinical pain monitoring.