Deep Learning-Driven Automatic Epileptic Seizure Detection Utilizing Stacked Autoencoders

Aryan     |     Nishchay     |     Mrs. Shanu

Chandigarh University, Mohali, India

aryanthakur1240@gmail.com  |  nishchaypabby@gmail.com  |  shanu.e19663@cumail.in

ABSTRACT

Epileptic seizure monitoring using scalp electroencephalography (EEG) remains a critical clinical challenge owing to the high dimensionality of EEG signals, substantial inter-patient variability, and the scarcity of annotated seizure events. This paper proposes a two-stage unsupervised detection framework — a Stacked Convolutional Autoencoder (SCAE) coupled with an LSTM Autoencoder (LSTM-AE) — trained exclusively on healthy (non-seizing) EEG recordings under a Leave-One-Patient-Out (LOPO) cross-validation protocol. The SCAE learns compact 128-dimensional spectral embeddings from multichannel EEG spectrograms; the LSTM-AE then captures temporal dynamics over sequences of these embeddings. Seizure episodes are identified as anomalies via elevated combined reconstruction error, with an adaptive per-patient threshold derived from the 99.95th percentile of validation scores. Post-processing employs a 70% vote-fraction smoothing window, 5-minute gap merging, and a minimum alarm-duration filter to suppress false positives. Evaluated on the CHB-MIT scalp EEG dataset (9 patients), the framework achieves a mean AUC of 0.89 (0.98 excluding the failure case), perfect sensitivity (1.0) for 6 of 9 patients, and an event-level false positive rate below 1 per hour for the best-performing patients. These results confirm that purely unsupervised hierarchical feature learning can deliver accurate, scalable seizure detection without any labelled seizure data.

Keywords: Epileptic Seizure Detection, EEG, Stacked Convolutional Autoencoder, LSTM Autoencoder, Anomaly Detection, Reconstruction Error, Zero-Shot Learning, CHB-MIT, Biomedical Signal Processing, LOPO Cross-Validation