Abstracts

Rationale: In few months after undergoing epilepsy surgery, 20–70% patients may start experiencing seizure recurrence, suggesting incomplete localization of the epileptogenic zone (EZ). The main challenge in accurately localizing the EZ is that each EZ is a unique and complex network, comprised of overlapping epileptogenic substrates that collectively cause seizures. Recent studies support the multi-component hypothesis for the EZ, demonstrating that combining biomarkers improves the EZ-localization compared to using individual biomarkers alone. Nonetheless, the feature high dimensionality (D) leads to challenges when training machine learning models for automated EZ-localization, limiting clinical utility.


Methods: We hypothesized that, although the EZ could be better characterized by high-D features, epileptogenicity, as a physiological construct, ought to have a low-D embedding. Operationally, it should be quantified by a continuum spanning from low to high seizure-risk, defined in a low-D latent space derived from high-D raw electrophysiological features. We tested this hypothesis in 3 steps. (I): Hundreds of raw features were extracted from 10-min of interictal stereo-EEG (SEEG), which were reduced to 10 eigen-features capable of differentiating clinically localized seizure zone (SZ). Two unsupervised classifiers were then trained to identify SZ in the eigen-feature space, which was used to fit a population level seizure-risk model. (II): the resting-SEEG trained risk model was next crossed-domain tested with 7–9 hours of sleep-SEEG in three patients from another cohort. (III): the cross-domain risk assessments were lastly validated using within-patient tensor component analysis (TCA) of the sleep-SEEG raw features.


Results: Across a broad parameter space, the unsupervised classifiers converged onto a consensus seizure-risk mode, characterized as a well-defined manifold in the eigen-feature space. For the resting-SEEG subjects, classification accuracy was determined by individual position in this eigen-feature space. The trained risk model exhibited robust cross-domain validity when characterizing time-varying seizure risk in sleep-SEEG. The individual TCA corroborated with the spatiotemporal characteristics of seizure-risk identified by the population level risk model.

Conclusions: These results provide compelling evidence to support our hypothesis that epileptogenicity has a low-D embedding, likely independent of brain states, pathological substrates, or other hidden variables. Our novel approach has minimal overhead, allowing for the integration of more biomarkers on large datasets, making it feasible for clinical use.

Funding:
The Academy of Finland (SP)
The Academy of Finland (JMP)?
Sigrid Jusélius Foundation, Finland (SHW)
The DARLING project (Projet-ANR-19-CE48-0002), Agence nationale de la recherche, France (MM)