Machine Learning for Localizing Epileptogenic-Zone in the Temporal Lobe: Quantifying the Value of Multimodal...

Background: Epilepsy affects 50 million people worldwide and a third are refractory to medication. If a discrete cerebral focus or network can be identified, neurosurgical resection can be curative. Most excisions are in the temporal-lobe, and are more likely to result in seizure-freedom than extra-temporal resections. However, less than half of patients undergoing surgery become entirely seizure-free. Localizing the epileptogenic-zone and individualized outcome predictions are difficult, requiring detailed evaluations at specialist centers. Methods: We used bespoke natural language processing to text-mine 3,800 electronic health records, from 309 epilepsy surgery patients, evaluated over a decade, of whom 126 remained entirely seizure-free. We investigated the diagnostic performances of machine learning models using set-of-semiology (SoS) with and without hippocampal sclerosis (HS) on MRI as features, using STARD criteria. Findings: Support Vector Classifiers (SVC) and Gradient Boosted (GB) decision trees were the best performing algorithms for temporal-lobe epileptogenic zone localization (cross-validated Matthews correlation coefficient (MCC) SVC 0.73 ± 0.25, balanced accuracy 0.81 ± 0.14, AUC 0.95 ± 0.05). Models that only used seizure semiology were not always better than internal benchmarks. The combination of multimodal features, however, enhanced performance metrics including MCC and normalized mutual information (NMI) compared to either alone (p < 0.0001). This combination of semiology and HS on MRI increased both cross-validated MCC and NMI by over 25% (NMI, SVC SoS: 0.35 ± 0.28 vs. SVC SoS+HS: 0.61 ± 0.27). Interpretation: Machine learning models using only the set of seizure semiology (SoS) cannot unequivocally perform better than benchmarks in temporal epileptogenic-zone localization. However, the combination of SoS with an imaging feature (HS) enhance epileptogenic lobe localization. We quantified this added NMI value to be 25% in absolute terms. Despite good performance in localization, no model was able to predict seizure-freedom better than benchmarks. The methods used are widely applicable, and the performance enhancements by combining other clinical, imaging and neurophysiological features could be similarly quantified. Multicenter studies are required to confirm generalizability. Funding: Wellcome/EPSRC Center for Interventional and Surgical Sciences (WEISS) (203145Z/16/Z).