Abstracts

Rationale: Epilepsy affects 50 million people worldwide, a third continue to have seizures despite medications. Surgery can cure epilepsy if a seizure-focus is identified. But less than half of surgeries result in seizure-freedom (Bell GS et al. J Neurol Neurosurg Psychiatry 2017; 88(11): 933-40).

Epileptic symptoms and signs help to localise the seizure focus in the evaluation of patients with drug resistant focal epilepsy for curative surgery, but there are relatively few clinical experts that can interpret these seizure manifestations. The abilities of initial semiologies to predict seizure foci are also variable and their variances have not been captured in large datasets (Luders HO. Textbook of epilepsy surgery: CRC Press; 2008. Elwan S et al. Seizure 2018; 61: 203-8)

We created the largest database relating seizure manifestations to their brain sources and tested how accurately we could predict seizure foci compared to the best clinical experts.

Methods: We created the Semiology-to-Brain Database from a PRISMA-guided individual-participant systematic review, integrated as Semiology Visualisation Tool (SVT) in the 3D-Slicer platform to objectively localise patients' seizure foci. The database yielded 11230 localising and 2391 lateralising semiology datapoints from 4643 patients across 309 studies. We mapped localisations to 105 geodesic information flow (GIF) brain parcellations and modelled them as binomial random variables.

We used SVT to retrospectively predict seizure foci in 14 randomly selected postsurgical adult patients (external to the database) who had epilepsy surgery and remained entirely seizure-free.

SVT evaluated the symptoms and signs in these 14 patients, using a Bayesian approach and inverse variance averaging for each brain region, to display predictions as probabilistic cortical heatmaps. These predictions were scored relative to the actual resected lobes and side of surgery. Finally, we compared SVT results to the scores of four expert epileptologists given the same clinical data.

Results: SVT was as good as expert clinicians in localising seizure foci (both scored 11/14 at best) but was even better at lateralising (+9 for SVT vs +7 for the best clinician).

Bayesian inference significantly enhanced SVT scores to match that of actual resections (Figs. 1 & 2).

Conclusions: SVT translates clinical phenotypes to anatomical imaging, is fully open-source, data-driven, and allows filtering by age using a graphical user interface.

We demonstrate that SVT predictions, despite being blinded to MRI and EEG data, are broadly congruent with actual resections for patients who are cured of their epilepsy after surgery, and as good as the best expert epileptologists’ predictions.

SVT could therefore be used as the basis for multimodal clinical decision support models to predict seizure foci in the presurgical evaluation of adult patients with drug-resistant focal epilepsy, with further validation required for use in children due to differing semiology (Ray A, Kotagal P. Epileptic Disord 2005; 7(4): 299-307).

Funding: Please list any funding that was received in support of this abstract.: Wellcome/EPSRC Centre for Interventional and Surgical Sciences (WEISS) (203145Z/16/Z).