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

Feng Jin ¹ Yunde Li ² ZiJun Huang ³ Yehang Chen ² SenLiang Lu ³ RongLiang Hu ⁴ Qinghui Hu ² YuYao Chen ⁵ Ximiao Wang ^4* Yong Fan ^2* He jing ^2*

• ¹ Guilin Normal College, Guilin, Guangxi Zhuang Region, China
• ² Guilin University of Aerospace Technology, Guilin, China
• ³ Guilin University of Electronic Technology, Guilin, Guangxi Zhuang Region, China
• ⁴ Jiangmen Central Hospital, Jiangmen, Guangdong, China
• ⁵ Nanxishan Hospital of Guangxi Zhuang Autonomous Region, Gulin, Guangxi Zhuang Region, China

Motor imagery functional near-infrared spectroscopy (MI-fNIRS) enables precise monitoring of neural activity in specific brain regions, offering significant benefits for stroke patients undergoing active-passive rehabilitation therapy. Accurate classification of MI-fNIRS signals in stroke patients holds critical clinical value. However, the limited availability of training samples due to the challenges associated with data collection and annotation, coupled with substantial inter-subject variability, results in poor generalization performance of cross-subject classification models. To address these challenges, this study proposes a Cross-Subject Heterogeneous Transfer Learning Model (CHTLM). The model leverages a large corpus of labeled electroencephalogram (EEG) data from healthy individuals performing different target tasks as the source domain. It employs an adaptive feature matching network to determine the alignment weights between feature maps and convolutional layers of both the source and target models, thereby enabling the transfer of task-relevant knowledge to corresponding components in the target model. Additionally, the model extracts multi-scale fNIRS features from the target domain and utilizes a sparse Bayesian extreme learning machine to classify the selected deep learning features.In this study, two MI-fNIRS datasets were collected from eight stroke patients before and after rehabilitation training for experimental evaluation. The results demonstrate that CHTLM achieves average classification accuracies of 0.831 and 0.913, with mean AUCs of 0.887 and 0.930, respectively. Compared to five baseline methods, CHTLM improves classification accuracy by 8.6%, 10.5%, 8.0%, 8.6%, and 10.5% before rehabilitation and by 11.3%, 15.7%, 13.2%, 11.3%, and 15.7% after rehabilitation.