Authors :
Presenting Author: Himanshu Kumar, PhD – Cleveland Clinic
David Martinez, MD – Cleveland Clinic
Jason Chisholm, MD – Cleveland Clinic
Jean Khoury, MD – Cleveland Clinic
Demitre Serletis, MD, PhD – Cleveland Clinic Epilepsy Center, USA
Imad Najm, MD – Cleveland Clinic
Andreas Alexopoulos, MD, MPH – Cleveland Clinic
Juan C. Bulacio, MD – Cleveland Clinic, Cleveland, United States
Balu Krishnan, PhD – Cleveland Clinic
Rationale:
Accurate identification of seizure onset patterns (SOPs) is crucial for understanding seizure dynamics and guiding surgical intervention in drug-resistant focal epilepsy [1]. Conventional SOP classification relies on expert interpretation and predefined features, which may overlook subtle or novel temporal-frequency patterns. Although machine learning has advanced seizure detection using SEEG [2], most methods are supervised and annotation-dependent. To address this limitation, we propose an unsupervised framework that leverages contrastive learning to derive latent embeddings from time-frequency representations of SEEG during seizure onset. This enables the discovery of meaningful ictal morphologies directly from data without expert labeling or prior assumptions, as supported by recent advances in unsupervised EEG analysis [3].
Methods:
SEEG data from 50 patients who underwent surgical resection were analyzed. For each seizure, 60-second peri-ictal segments (including 20 seconds preictal) were transformed into Morlet-based time-frequency maps, yielding 27,568 images. These were used to train the Seizure Onset Pattern Encoder Network (SOPENet)—a deep model with convolutional layers, a transformer encoder, attention pooling, and a contrastive projection head. The model produced 128-dimensional embeddings capturing spectro-temporal dynamics. Clustering focused on SEEG data from epileptic foci of Engel I patients. The optimal cluster number (k = 7) was selected using standard metrics. Final clustering combined K-means with t-distribution refinement, and clusters were visualized using t-SNE.
Results:
Clustering of 385 seizure segments revealed seven distinct ictal morphologies:
-
C1: Low-frequency preictal transitioning into 14–75 Hz fast activity.
-
C2: Preictal spiking and high-frequency onset with low-frequency suppression.
-
C3: Beta evolving into gamma ( >75 Hz) with marked low-frequency attenuation.
-
C4: Preictal spikes followed by broad suppression at onset.
-
C5: Abrupt broadband onset without preictal spikes, evolving into rhythmic bursts.
-
C6: Preictal spikes with short-duration broadband activity.
-
C7: Persistent repetitive spiking with minimal evolution.
t-SNE visualization showed strong intra-cluster cohesion and inter-cluster separation, highlighting the discriminative power of the learned embeddings and the utility of unsupervised methods for SOP discovery [3].
Conclusions:
Our unsupervised contrastive learning framework effectively captures latent ictal onset representations from SEEG, enabling objective identification of seizure morphologies without expert labeling. This supports data-driven seizure characterization and complements efforts in machine learning for epileptogenic zone localization [2], paving the way for scalable, personalized epilepsy analysis using high-resolution SEEG data.
References:
[1] Lagarde S, et al. Epilepsia. 2019;60:85–95.
[2] Jose B, et al. Brain Res. 2023;1820:148546.
[3] Yıldız İ, et al. Comput Methods Programs Biomed. 2022;215:106604.
Funding: NA