Construction of a deep - learning - based rehabilitation prediction model for lower-limb motor dysfunction after stroke using synchronous EEG-EMG and fMRI

Objective: Construct a predictive model for rehabilitation outcomes in ischemic stroke patients 3 months post-stroke using resting state functional magnetic resonance imaging (fMRI) images, as well as synchronized electroencephalography (EEG) and electromyography (EMG) time series data. Methods: A total of 102 hemiplegic patients with ischemic stroke were recruited. Resting - state functional magnetic resonance imaging (fMRI) scans were carried out on all patients and 86 of them underwent simultaneous electroencephalogram (EEG) and electromyogram (EMG) examinations. After data preprocessing, we established prediction models based on time-series data and fMRI images separately. The predictions of the time - series model and the fMRI model were integrated using ensemble learning methods to create a multimodal fusion prediction model. The accuracy, recall, precision, F1 - score, and the area under the ROC curve (AUC) were calculated to evaluate the performance of the model. Results: Compared to unimodal prediction models, multimodal fusion models demonstrated superior predictive performance. The ShuffleNet-LSTM model outperformed other multimodal fusion approaches. The area under the ROC curve was 0.8665, accuracy was 0.8031, F1-score was 0.7829, recall was 0.774, and precision was 0.833. Conclusion: A deep learning-based rehabilitation prediction model utilizing multimodal signals was successfully developed. The ShuffleNet-LSTM model exhibited excellent performance among multimodal fusion models, effectively enhancing the accuracy of predicting lower-limb motor function recovery in stroke patients.