Variational autoencoder for explainable seizure onset phases detection

Abstract Objective We present the first deep learning framework for automated, time-resolved, per-channel annotation of ictal and Low-Voltage Fast Activity (LVFA) onsets in stereo electroencephalography (SEEG) recordings of patients with focal epilepsy. To our knowledge, no prior system jointly addresses these tasks on continuous single-channel recordings. Approach A one-dimensional Variational Autoencoder (VAE) encodes 2-second SEEG segments into a 60-dimensional latent space and classifies them as interictal, ictal, or LVFA via a linear classifier. A postprocessing algorithm converts segment-level probabilities into per-channel onset markers at 0.5-second resolution. The system was trained and evaluated using subject-wise 5-fold cross-validation on 37 patients with manual ictal and LVFA annotations. Main results At the segment level, the VAE classified the three classes with an average recall of 0.88. At the channel level, it reached an ictal recall of 0.84 (0.91 on Seizure Onset Zone channels) and LVFA recall of 0.74, with median onset latencies of 5.0 s and 0.86 s, respectively. As a seizure detector, the system achieved 99.1 % recall with 1 % false positives. Latent dimensions correlated with physiologically interpretable features (amplitude, band powers, spectral flatness, energy ratio). An ablation study showed that the VAE’s reconstruction objective provides dual benefits over a discriminative encoder-only baseline: improved detection performance and stronger alignment between latent dimensions and these clinically meaningful features. Significance By providing the first time-resolved per-channel framework for joint ictal and LVFA annotation, this work establishes a robust and explainable platform for automated SEEG analysis with potential to substantially reduce clinician workload during presurgical epilepsy evaluation.