bioRxiv2025
Accurately localizing the epileptogenic network (EpiNet) remains a major barrier to effective epilepsy treatment, largely due to limited mechanistic understanding. The EpiNet is a patient-specific brain network shaped by complex, overlapping pathology. While combining biomarkers can improve localization, it also generates high-dimensional feature data that increases the risk of overfitting and reduces interpretability. We hypothesized that the core epileptogenic dynamics could be captured in a low-dimensional latent space derived from empirical data, without the need to record seizures. From interictal stereo-EEG (SEEG) recordings in 64 patients (29 females), we extracted 260 neuronal features and reduced them to 10 latent components using singular value decomposition. A classifier trained on these 10 components was then simplified into a probabilistic EpiNet model requiring only two components as input. Individual position in this two-dimensional latent space correlated with previously reported classification accuracy (r2=0.5), supporting its functional relevance. In three independent patients, the probabilistic model captured time-varying epileptogenic dynamics during sleep-SEEG recordings, corroborated clinical assessments, and achieved peak classification accuracies of 0.63, 0.85, and 0.94. These predictions were independently validated by tensor component analysis. Together, these results provide evidence for a robust low-dimensional representation of epileptogenicity across brain states and pathological substrates. This approach simplifies interpretation, facilitates integration of additional biomarkers, and enables large-scale cohort analyses, establishing a proof of concept for a unified framework for epilepsy biomarkers. Significance StatementTo advance mechanistic understanding of large-scale brain dynamics underlying epilepsy, we combined novel epilepsy biomarkers with interpretable machine learning. From interictal SEEG, we extracted 260 connectivity and criticality features. Dimensionality reduction of these raw features showed that only two components were needed to identify epileptogenic networks, reducing the feature space by >99%. These components were highlighted by abnormal power-law scaling, bistability, and strong inhibition or excitation in the 15-200 Hz range, consistent with prior findings. The model tracked neuropathological dynamics over hours and was validated through tensor component analysis, suggesting that epileptogenic activity is dynamic, subject-specific, and sparsely represented in both state space and cortical networks.