A novel way to use cross-validation to measure connectivity by machine learning allows epilepsy surgery outcome prediction.

The rate of success of epilepsy surgery, ensuring seizure-freedom, is limited by the lack of epileptogenicity biomarkers. Previous evidence supports the critical role of functional connectivity during seizure generation to characterize the epileptogenic network (EN). However, EN dynamics is highly variable across patients, hindering the development of diagnostic biomarkers. Without relying on specific connectivity variables, we focused on a general hypothesis that the EN undergoes the greatest magnitude of connectivity change during seizure generation, compared to other brain networks. To test this hypothesis, we developed a novel method for quantifying connectivity change between network states and applied it to identify surgical resection areas. A network state was represented by random snapshots of connectivity within a defined time interval of an intracranial EEG recording. A binary classifier was applied to classify two network states. The classifier generalization performance estimated by cross-validation was employed as a continuous measure of connectivity change. The algorithm generated a network by iteratively adding nodes until the connectivity change magnitude decreased. The resulting network was compared to the surgical resection, and the overlap score was used to predict post-surgical outcomes. The framework was evaluated in a consecutive cohort of 21 patients with a post-surgical follow-up of minimum 3 years. The best overlap between connectivity change networks and resections was obtained at the transition from pre-seizure to seizure (surgical outcome prediction ROC-AUC=90.3 %). However, all patients except one were correctly classified when considering the most informative time intervals. Time intervals proportional to seizure length were more informative than the almost universally used fixed intervals. This study demonstrates that connectivity can be successfully classified with a machine learning analysis and provide information for distinguishing a separate epileptogenic functional network. In summary, the connectivity change analysis could accurately identify epileptogenic networks validated by surgery outcome classification. Connectivity change magnitude at seizure transition could potentially serve as an EN biomarker. The tool provided by this study may aid surgical decision-making.