Unsupervised clustering of electroclinical features in temporal lobe epilepsy: A data-driven approach

To identify clinically meaningful patterns in ictal electroclinical features of focal epilepsy using a data-driven, unsupervised learning approach, and to assess whether such patterns can localize and lateralize the epileptogenic zone (EZ) more accurately than conventional electroclinical interpretation.

Ictal scalp EEG and seizure semiology were systematically extracted using the SCORE system in 116 patients with drug-resistant focal epilepsy who subsequently underwent resective surgery. Unsupervised clustering was applied separately to EEG and semiology features using k-means and agglomerative hierarchical clustering. Cluster quality was evaluated using the silhouette score, within-cluster sum of squares (WSS), between-cluster sum of squares (BSS), the ratio of between-cluster to within-cluster variance (BSS/WSS), the proportion of variance explained by clustering (BSS/TSS), and the Calinski-Harabasz index. The ability of the resulting clusters to localize and lateralize the temporal lobe EZ was assessed and compared with the conventional approach reflecting current clinical practice, based on previously published lateralizing signs, semiology localizing to the temporal lobe, and scalp EEG localization of ictal onset.

Unsupervised clustering achieved strong structure for EEG and moderate for semiology data. K-means clustering of EEG features yielded a silhouette score of .63, while hierarchical clustering of semiology features achieved a silhouette score of .54. Semiology-based clusters showed substantial overlap with respect to both localization and lateralization. EEG-based clusters localized and lateralized the temporal lobe EZ with an accuracy of 65.3% (95% CI: 55.0%-74.6%). This performance was comparable to conventional electroclinical interpretation using scalp EEG ictal onset (68.1%, 95% CI: 57.5%-77.5%). Interpretation based on seizure semiology achieved an accuracy of 71.3% (95% CI: 61.0-80.1%).

Unsupervised clustering of ictal electroclinical features did not outperform conventional expert interpretation in localizing or lateralizing the EZ. These findings suggest that the limitations may be inherent to the observable ictal phenomena captured by scalp EEG and semiology. Future research should focus on identifying novel or complementary biomarkers to improve precision in epilepsy surgery evaluation.