Variational autoencoder for interpretable seizure onset phases detection.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Informatics

Background:

• Focal epilepsy diagnosis relies on interpreting stereo electroencephalography (SEEG) data.
• Manual annotation of SEEG data is time-consuming and requires specialized expertise.
• Automated methods are needed to improve the efficiency of SEEG data analysis.

Purpose of the Study:

• To develop and evaluate a deep learning framework for automated seizure annotation in SEEG data.
• To utilize a Variational Autoencoder (VAE) for feature extraction and classification of SEEG segments.
• To assess the system's performance in detecting seizures and Low-Voltage Fast Activity (LVFA).

Main Methods:

• A deep learning framework combining a 1D VAE for feature extraction and a linear classifier for segment classification was developed.
• The 1D VAE encodes 2-second SEEG segments into a low-dimensional latent space.
• The system was trained and validated using SEEG data from 37 epilepsy patients with clinician-provided annotations, employing 5-fold cross-validation.

Main Results:

• The system achieved an average recall of 0.88 for classifying ictal vs. interictal 2-second segments.
• Whole-channel seizure annotation demonstrated high average recall (0.86 for ictal/LVFA, 0.91 on Seizure Onset Zone channels).
• Temporal accuracy for seizure onset and LVFA detection was high, with mean time lags of 9.8 and 2.0 seconds, respectively. Seizure detection recall reached 99% with 95% specificity.

Conclusions:

• A VAE-based deep learning approach effectively generates a meaningful latent space from SEEG data.
• This framework shows potential for accurate seizure and fast onset pattern detection in epilepsy.
• The developed system may significantly reduce the clinical workload associated with SEEG data review.