Variational autoencoder for interpretable seizure onset phases detection

In this study, we describe a deep learning framework for automated seizure annotation in stereo electroencephalography (SEEG) data of patients with focal epilepsy. We use a one-dimensional Variational Autoencoder (VAE) for feature extraction of single-channel temporal series and a linear classifier for segment classification. We trained the network using data from 37 patients containing manual annotations of ictal and Low-Voltage Fast Activity (LVFA) by clinicians. The 1D VAE encodes two-second SEEG segments into a low-dimensional representation in latent space and then classifies them as interictal, ictal, or LVFA segments. We used 5-fold cross-validation for training and validation. Our system classified ictal vs. interictal 2 second segments with an average recall of 0.88. For whole-channel seizure annotation, we combine the probabilities of all its segments and compute the Area Under Curve (AUC) of the exponentially smoothed probability signal, marking the onset of both ictal and LVFA, a high average recall of 0.86, with even higher performance (0.91) on channels identified by clinicians as belonging to the Seizure Onset Zone (SOZ). The markers were also temporally accurate, with a mean (median) time lag of 9.8 (5.0) seconds for the ictal onset and a mean (median) lag of 2.0 (0.8) seconds for LVFA. Additionally, latent space analysis suggests that multiple dimensions correlate with class-relevant SEEG features that the classifier pays attention to, most notably amplitude and spectral power, which provides an explainability component to the network. As a secondary objective, we obtain a seizure detection recall of 99% with a specificity of 95%. Our findings suggest that a VAE-based approach can produce a meaningful latent space from SEEG data and leverage it to detect seizures and fast onset patterns, potentially helping clinicians reduce the workload of SEEG review.