TPCNet: A Temporal Periodicity Convolutional Network for motor imagery EEG decoding in stroke patients

Stroke caused by vascular rupture or blockage has high incidence and leads to significant disability. Motor imagery (MI) electroencephalogram (EEG) is a promising approach to understanding and addressing stroke-related motor impairments. However, the practical application of EEG-based rehabilitation is hindered by an insufficient understanding of the task-specific features and complex temporal patterns inherent in the EEG signals of stroke patients.

In this study, we collected EEG signals from 24 stroke patients performing four unilateral upper limb MI tasks. Among them, 12 subjects performed forward arm raising and lowering, while the remaining 12 performed lateral arm raising and lowering. Moreover, we propose a Temporal Periodicity Convolutional Network (TPCNet) for EEG-based MI classification. TPCNet consists of a convolutional block for extracting shallow spatiotemporal features, a sliding window structure that ensures consistent action initiation across samples, and a temporal periodicity block for capturing variations in periodic patterns associated with MI tasks.

TPCNet achieved a classification accuracy of 86.53% on the stroke patient MI dataset and 82.21% on the BCI Competition IV 2a dataset (left hand, right hand, feet, and tongue). Gradient-weighted Class Activation Mapping (Grad-CAM) analysis suggests that stroke patients may exhibit longer task-specific MI periodicity than healthy subjects.

The proposed method achieves superior performance on stroke patient MI tasks and competitive results on public MI datasets involving healthy subjects.

The proposed TPCNet model effectively captures the spatiotemporal features and periodic patterns of EEG signals, leading to enhanced classification accuracy.