Adaptive Temporal-Scale Mamba with Rhythm-Aware Gating for the Classification of Ictal, Interictal and Normal EEG Segments

This is the authors' abstract. We don't add key points for this paper.

Electroencephalogram (EEG)-based classification of ictal, interictal, and normal segments is a core building block of automated seizure-monitoring systems. Existing deep models still struggle to jointly optimize sensitivity, specificity, cross-subject generalization, and computational cost. Motivated by the multi-rhythm structure of EEG, this study proposes an Adaptive Temporal-Scale Mamba with Rhythm-Aware Gating (ATSM-RAG) model. A Rhythm-Aware Gating (RAG) unit maps the relative band power of each EEG segment to a scalar gate that interpolates the discretization step (the effective temporal scale) of a Mamba state-space backbone, adapting the temporal receptive field to the input’s spectral state rather than relying on a fixed time constant. ATSM-RAG is evaluated on two public corpora under patient-independent protocols, and we foreground the two evaluations that are free of cross-corpus confounds. First, on the confound-free binary CHB-MIT benchmark (ictal versus interictal, patient-level leave-one-patient-out cross-validation, 22 folds), ATSM-RAG achieves an accuracy of 99.62% ± 0.35%, a sensitivity of 99.10% ± 0.95%, and a specificity of 99.75% ± 0.30%. Second, following the SzCORE recommendations, event-based scoring on all common-montage CHB-MIT recordings (633 files, 930.6 hours, 165 annotated seizures) yields an event-level sensitivity of 90.3% at 2.4 false alarms per 24 hours, with a median detection latency of 12.3 seconds. For completeness, the ternary benchmark (288 ictal and 1,440 interictal CHB-MIT segments plus 720 normal segments drawn from clinically normal recordings of the TUH Abnormal EEG Corpus) yields an accuracy of 99.51% ± 0.37%, a sensitivity (ictal recall) of 98.99% ± 1.29%, and a specificity (correct-classification rate of the merged non-ictal classes) of 99.58% ± 0.39%. On the TUH EEG Seizure Corpus (official patient-disjoint train/development/evaluation split), ATSM-RAG achieves a binary segment classification accuracy of 97.94%. Ablation experiments show that removing the RAG unit and fixing the temporal scale reduce accuracy by 2.94% and 2.05%, respectively. These results indicate that rhythm-conditioned temporal scaling is a promising and computationally efficient inductive bias for EEG segment classification; prospective and multi-center clinical validation is still required before any clinical deployment.