A Novel Transfer Learning-Based Hybrid EEG-fNIRS Brain-Computer Interface for Intracerebral Hemorrhage Rehabilitation

Abstract

Motor imagery (MI)-based neurorehabilitation shows promise for intracerebral hemorrhage (ICH) recovery, yet conventional unimodal brain-computer interfaces (BCIs) face critical limitations in cross-subject generalization. This study presents a multimodal electroencephalography (EEG)-functional near-infrared spectroscopy (fNIRS) fusion framework incorporating a Wasserstein metric-driven source domain selection method that quantifies inter-subject neural distribution divergence. Through comparative neuroactivation analysis of 17 normal controls and 13 ICH patients during MI tasks, the transfer learning model achieved 74.87% mean classification accuracy on patient data when trained with optimally selected normal templates. Cross-validation on two public hybrid EEG-fNIRS datasets demonstrated generalizability, increasing baseline accuracy to 82.30% and 87.24%, respectively. The proposed system synergistically combines the millisecond temporal resolution of EEG with the hemodynamic spatial specificity of fNIRS, establishing the first clinically viable multimodal analytical protocol for ICH rehabilitation. This paradigm advances neurotechnology translation by paving the way for personalized rehabilitation regimens through robust cross-subject neural pattern transfer while addressing the critical barrier of neurophysiological heterogeneity in post-ICH populations.

Keywords: brain‐computer interfaces; electroencephalograph; functional near infrared imaging; intracerebral hemorrhage; motor imagery; transfer learning.

Figure 1

Overview of the study design.

Figure 2

Time‐frequency domain analysis of EEG signals for the normal subject A) and the patient with intracerebral hemorrhage B).

Figure 3

A) Inter‐subject Wasserstein distance heatmap for EEG modality in Dataset_N; B) Inter‐subject Wasserstein distance heatmap for fNIRS modality in Dataset_N; C) Inter‐subject Wasserstein distance heatmap for EEG modality in Dataset_P; D) Inter‐subject Wasserstein distance heatmap for fNIRS modality in Dataset_P; E) Optimal source template for Dataset_N (EEG); F) Optimal source template for Dataset_N (fNIRS); G) Optimal source template for Dataset_P (EEG); H) Optimal source template for Dataset_P (fNIRS); I) Radar plot of motor imagery classification accuracy for each subject; J) Mean classification accuracy in Dataset_N and Dataset_P across modalities and methods.

Figure 4

A) Optimal source template for Dataset_N (EEG + fNIRS); B) Optimal source template for Dataset_P (EEG + fNIRS); C) Confusion Matrix of Dataset_N; D) Confusion Matrix of Dataset_P.

Figure 5

A) Inter‐subject Wasserstein distance heatmap for EEG modality in Dataset_MI; B) Inter‐subject Wasserstein distance heatmap for fNIRS modality in Dataset_MI; C) Optimal source template for the Dataset_MI (EEG and fNIRS); D) Radar plot of motor imagery classification accuracy for each subject; E) Performance comparison between the proposed method and related studies. Pth‐PF: pth‐order polynomial fusion; FWHTC _svd : the singular value decomposition values of the Fast Walsh‐Hadamard transform coefficients; KNN: k‐nearest neighbor algorithm.

Figure 6

A) Inter‐subject Wasserstein distance heatmap for EEG modality in Dataset_MA; B) Inter‐subject Wasserstein distance heatmap for fNIRS modality in Dataset_MA; C) Optimal source template for the Dataset_MA (EEG and fNIRS); D) Radar plot of motor imagery classification accuracy for each subject; E) Performance comparison between the proposed method and related studies.

Figure 7

A) Optimal source template for Dataset_MI (EEG + fNIRS); B) Optimal source template for Dataset_MA (EEG + fNIRS); C) Confusion Matrix of Dataset_MI; D) Confusion Matrix of Dataset_MA.

Figure 8

A) Channel configuration and experimental paradigm for the private dataset; B) Channel configuration and experimental paradigm for the public motor imagery dataset; C) Schematic diagram of depthwise separable convolution; D) Architecture of the multimodal fusion model.