The Value of Seizure Semiology in Epilepsy Surgery: Epileptogenic-Zone Localisation in Presurgical Patients using Machine Learning and Semiology Visualisation Tool

Background Eight million individuals have focal drug resistant epilepsy worldwide. If their epileptogenic focus is identified and resected, they may become seizure-free and experience significant improvements in quality of life. However, seizure-freedom occurs in less than half of surgical resections. Seizure semiology - the signs and symptoms during a seizure - along with brain imaging and electroencephalography (EEG) are amongst the mainstays of seizure localisation. Although there have been advances in algorithmic identification of abnormalities on EEG and imaging, semiological analysis has remained more subjective. The primary objective of this research was to investigate the localising value of clinician-identified semiology, and secondarily to improve personalised prognostication for epilepsy surgery. Methods I data mined retrospective hospital records to link semiology to outcomes. I trained machine learning models to predict temporal lobe epilepsy (TLE) and determine the value of semiology compared to a benchmark of hippocampal sclerosis (HS). Due to the hospital dataset being relatively small, we also collected data from a systematic review of the literature to curate an open-access Semio2Brain database. We built the Semiology-to-Brain Visualisation Tool (SVT) on this database and retrospectively validated SVT in two separate groups of randomly selected patients and individuals with frontal lobe epilepsy. Separately, a systematic review of multimodal prognostic features of epilepsy surgery was undertaken. The concept of a semiological connectome was devised and compared to structural connectivity to investigate probabilistic propagation and semiology generation. Results Although a (non-chronological) list of patients’ semiologies did not improve localisation beyond the initial semiology, the list of semiology added value when combined with an imaging feature. The absolute added value of semiology in a support vector classifier in diagnosing TLE, compared to HS, was 25%. Semiology was however unable to predict postsurgical outcomes. To help future prognostic models, a list of essential multimodal prognostic features for epilepsy surgery were extracted from meta-analyses and a structural causal model proposed. Semio2Brain consists of over 13000 semiological datapoints from 4643 patients across 309 studies and uniquely enabled a Bayesian approach to localisation to mitigate TLE publication bias. SVT performed well in a retrospective validation, matching the best expert clinician’s localisation scores and exceeding them for lateralisation, and showed modest value in localisation in individuals with frontal lobe epilepsy (FLE). There was a significant correlation between the number of connecting fibres between brain regions and the seizure semiologies that can arise from these regions. Conclusions Semiology is valuable in localisation, but multimodal concordance is more valuable and highly prognostic. SVT could be suitable for use in multimodal models to predict the seizure focus