Efficient seizure detection by lightweight Informer combined with fusion of time–frequency–spatial features

Automatic seizure detection based on electroencephalogram (EEG) signals is essential for monitoring and diagnosing epilepsy, as well as reducing the workload of neurologists who visually inspect long-term EEGs. In this work, a novel framework for automatic seizure detection is proposed by integrating the Stockwell transform (S-transform) with a lightweight Informer model. The S-transform is firstly used to convert EEG signals into multi-level time–frequency features. Subsequently, an Informer encoder is deployed to capture spatial and long-term dependencies of these EEG time–frequency features and perform classification for seizure detection. Both the segment-based evaluation and event-based evaluation were conducted on the CHB-MIT EEG database and the QH-SDU database in patient-specific scenarios. Due to the efficient multi-resolution time–frequency analysis capability of the S-transform and the Informer’s ability to measure spatio-temporal correlation with lower time complexity and memory usage, the proposed method achieved state-of-the-art outcomes over the two EEG databases. The experimental results substantiate the model's ability to generalize across different databases and potential for clinical application.