Machine learning models trained in a low-dimensional latent space for epileptogenic zone (EZ) localization

Identifying the epileptogenic zone (EZ) for epilepsy surgery is challenging because each patient’s EZ represents a unique complex network. This network may include overlapping pathological brain tissues exhibiting high-dimensional (D) features. Having high-D features, however, could lead to problems when training machine learning (ML) models for automated EZ-localization. We posited that, albeit exhibiting high-D features, epileptogenicity of the EZ, as a construct, ought to have a low-D embedding that can be defined by a gradient from low to high seizure risk. We proposed a two-step approach involving dimensionality reduction for feature-selection in a low-D latent space and subsequent training of unsupervised ML models for a consensus classification of clinically defined seizure-zone (SZ). We extracted hundreds of raw electrophysiological features from brain regions using interictal resting-state stereoelectroencephalography (SEEG). These raw features were then reduced to ten eigen-features capable of differentiating the SZ. We next trained two ML algorithms using these eigen-features to identify the SZ. Across a broad parameter space, the algorithms converged on a consensus seizure-risk mode in the eigen-feature space. This resting-SEEG derived risk model showed cross-domain validity for characterizing epileptogenicity in sleep-SEEG from a patient with different pathological substrates, thereby offering preliminary evidence to support our low-D epileptogenicity proposal.