AM-EEGNet: An advanced multi-input deep learning framework for classifying stroke patient EEG task states

Stroke is the leading cause of adult disability among all prevalent pathologies around the world. To improve post-stroke patients' active daily life and living quality, revealing the underlying brain mechanism of stroke recovery is crucial. The EEG feature signals (power spectrum density and functional connectivity) in two different states (eyes-close, eyes-open) show their ability as predictors in post-stroke recovery. In addition, deep learning methods can successfully extract EEG features to predict. To this end, we propose an advanced multi-input deep-learning framework that can extract multi-EEG feature signals and explain results from EEG feature inputs for stroke patients. A total of 72 post-stroke patients were recruited in this study. Each would be asked to participate in two experiments (eyes-closed and eyes-open resting state). The deep learning framework would be based on their EEG feature signals to predict their task states. AM-EEGNet achieves high performance (Accuracy: 97.22%, Sensitivity: 0.94, and Specificity: 1.00) in the EEG-based states classification problems. In addition, we demonstrated the explanation result from EEG features. Our results suggest that AM-EEGNet is robust enough to learn EEG features from stroke patients and can explain the EEG features related to tasks. Moreover, our results reveal the difference in those two eyes-close and eyes-open resting states for stroke patients. Model details can be found at https://github.com/linbingru/am-eegnet.