Heterogeneous transfer learning model for improving the classification performance of fNIRS signals in motor imagery among cross-subject stroke patients.

Introduction: Motor imagery functional near-infrared spectroscopy (MI-fNIRS) offers precise monitoring of neural activity in stroke rehabilitation, yet accurate cross-subject classification remains challenging due to limited training samples and significant inter-subject variability. This study proposes a Cross-Subject Heterogeneous Transfer Learning Model (CHTLM) to enhance the generalization of MI-fNIRS signal classification in stroke patients. Methods: CHTLM leverages labeled electroencephalogram (EEG) data from healthy individuals as the source domain. An adaptive feature matching network aligns task-relevant feature maps and convolutional layers between source (EEG) and target (fNIRS) domains. Multi-scale fNIRS features are extracted, and a sparse Bayesian extreme learning machine classifies the fused deep learning features. Results: Experiments utilized two MI-fNIRS datasets from eight stroke patients pre- and post-rehabilitation. CHTLM achieved average accuracies of 0.831 (pre-rehabilitation) and 0.913 (post-rehabilitation), with mean AUCs of 0.887 and 0.930, respectively. Compared to five baselines, CHTLM improved accuracy by 8.6-10.5% pre-rehabilitation and 11.3-15.7% post-rehabilitation. Discussion: The model demonstrates robust cross-subject generalization by transferring task-specific knowledge from heterogeneous EEG data while addressing domain discrepancies. Its performance gains post-rehabilitation suggest clinical potential for monitoring recovery progress. CHTLM advances MI-fNIRS-based brain-computer interfaces in stroke rehabilitation by mitigating data scarcity and variability challenges.