期刊: [Institute of Electrical and Electronics Engineers] 日期:2026-08-01卷期号:: 1-6
标识
DOI:10.1109/icsseecc69062.2026.11722036
摘要
Clinical deployment of motor imagery BCI systems is hindered by the non-stationary, pathological nature of post-stroke EEG and the predominant monomodal paradigm that ignores the patient’s physiological condition. To address this research gap, we propose two multimodal late-fusion architectures, MM-ATCNet and MM-EEGNetCBAM evaluated on the UET175 dataset of 30 acute stroke patients, representing the first deep learning benchmark on this cohort. Both architectures fuse spatial-temporal MI-EEG features with a clinical metadata vector comprising NIHSS, mRS, muscle strength, age, and sex. To disentangle architectural robustness from optimisation-dependent variation, models are assessed under a rigorous 3x3 grid of cross-validation folds (K=5,8,10) and optimizers (Adam, AdamW, RAdam). MM-ATCNet with AdamW achieves a peak accuracy of 65.33and AUC=0.69, outperforming the Riemannian TWFB+MDM baseline by 11.07 percentage points and unimodal EEGNet by 6.73 points. Results confirm that fusing structured clinical severity scores with EEG decoding yields consistent, individualized performance gains across all nine experimental configurations.