A Novel Deep Learning Model for Epileptic Seizure Detection from EEG Data

Abstract

Electroencephalogram (EEG) is critical for diagnosing epilepsy, a chronic neurological disorder affecting over 70 million people worldwide and characterized by abnormal neuronal discharges causing recurrent seizures. Manual EEG analysis is labor-intensive and struggles with large data volumes, necessitating efficient automated solutions. While existing seizure detection methods predominantly use supervised learning, they face challenges due to rare annotated ictal events and data imbalance. Emerging unsupervised approaches offer promise, such as transformer-based models for label-free EEG activity identification [1].

By framing seizure identification as an anomaly detection problem, we aim to propose an unsupervised anomaly detection framework using an xLSTM (Extended Long Short-Term Memory) Autoencoder to detect epileptic seizures without expert-labeled data. To our knowledge, this is the first application of xLSTM Autoencoders to unsupervised anomaly detection in time-series EEG. This architecture leverages xLSTM’s enhanced memory mechanisms, improving sequence modeling over traditional LSTMs [2].

The model will be trained exclusively on normal EEG segments from healthy subjects. Seizures will then be identified as anomalies via high reconstruction errors when input patterns deviate from learned normal distributions. The encoder (residual mLSTM and sLSTM blocks) compresses EEG segments into latent vectors, while the decoder (residual mLSTM and sLSTM blocks) reconstructs the sequence. Anomaly scoring is usually done by measuring reconstruction error and flagging segments above a threshold as seizures.

We use a curated subset of the Temple University Hospital TUH EEG Corpus, the TUH EEG Epilepsy Corpus (TUEP), containing recordings from 100 epilepsy patients and 100 healthy controls [3]. Two unipolar montages were used in the recordings: Average Reference (AR) and Linked Ears Reference (LE), of which only the LE montage is selected due to its stability and artifact robustness. Additionally, non-shared channels across recordings are excluded, retaining only common channels.

To allow the model to learn its own latent space from raw signals, we did not extract features. Instead, we applied a minimal preprocessing pipeline: a [0.5–49 Hz] bandpass filter to retain seizure-relevant frequencies and remove artifacts, z-score normalization and resampling to 250 Hz. We finally segmented the recordings into one-second windows with 50% overlap.

While the preprocessing pipeline is complete, model implementation and training on normal EEG segments are underway. We plan to evaluate performance using accuracy, precision, recall, F1-score, and AUC-ROC. Future steps include optimizing hyperparameters via evolutionary algorithms, benchmarking against state-of-the-art methods (e.g., [1]) and validating the model’s real-time detection capability.

References

[1] İ. Y. Potter, G. Zerveas, C. Eickhoff, and D. Duncan, “Unsupervised Multivariate Time-Series Transformers for Seizure Identification on EEG,” 2022 21st IEEE International Conference on Machine Learning and Applications (ICMLA), Nassau, Bahamas, 2022, pp. 1304-1311. doi: 10.1109/ICMLA55696.2022.00208.

[2] M. Beck et al., “xLSTM: Extended Long Short-Term Memory,” arXiv, 2024. doi: 10.48550/ARXIV.2405.04517.

[3] S. I. Choi, S. Lopez, I. Obeid, M. Jacobson, and J. Picone, “The Temple University Hospital EEG Corpus.” 2017. [Online]. Available: http://www.isip.piconepress.com/projects/tuh_eeg