GEM-CRAP: a fusion architecture for focal seizure detection

Abstract

Abstract Background Identification of seizures is essential for the treatment of epilepsy. Current machine-learning and deep-learning models often perform well on public datasets when classifying generalized seizures with prominent features. However, their performance was less effective in detecting brief, localized seizures. These seizure-like patterns can be masked by fixed brain rhythms. Methods Our study proposes a supervised multilayer hybrid model called GEM-CRAP (gradient-enhanced modulation with CNN-RES, attention-like, and pre-policy networks), with three parallel feature extraction channels: a CNN-RES module, an amplitude-aware channel with attention-like mechanisms, and an LSTM-based pre-policy layer integrated into the recurrent neural network. The model was trained on the Xuanwu Hospital and HUP iEEG dataset, including intracranial, cortical, and stereotactic EEG data from 83 patients, covering over 8500 labeled electrode channels for hybrid classification (wakefulness and sleep). A post-SVM network was used for secondary training on channels with classification accuracy below 80%. We introduced an average channel deviation rate metric to assess seizure detection accuracy. Results For public datasets, the model achieved over 97% accuracy for intracranial and cortical EEG sequences in patients, and over 95% for mixed sequences, with deviations below 5%. In the Xuanwu Hospital dataset, it maintained over 94% accuracy for wakefulness seizures and around 90% during sleep. SVM secondary training improved average channel accuracy by over 10%. Additionally, a strong positive correlation was found between channel accuracy distribution and the temporal distribution of seizure states. Conclusions GEM-CRAP enhances focal epilepsy detection through adaptive adjustments and attention mechanisms, achieving higher precision and robustness in complex signal environments. Beyond improving seizure interval detection, it excels in identifying and analyzing specific epileptic waveforms, such as high-frequency oscillations. This advancement may pave the way for more precise epilepsy diagnostics and provide a suitable artificial intelligence algorithm for closed-loop neurostimulation.