Machine Learning for Localising Epileptogenic-Zone in the Temporal Lobe: Quantifying the Value of Multimodal Clinical-Semiology and Imaging Concordance

Abstract

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 actually become entirely-seizure-free. Localising the epileptogenic-zone and individualised outcome predictions are difficult, requiring detailed evaluations at specialist centres. 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: Gradient boosted decision trees were one of the best performing algorithms for temporal-lobe epileptogenic zone localisation (cross-validated Matthews correlation coefficient (MCC) 0·64 ± 0·27, balanced accuracy 0·81 ± 0·14). Models using SoS features did not always improve scores above internal benchmarks. The combination of multimodal features enhanced performance metrics including MCC and normalised mutual information (NMI) compared to either alone (p<0·0001). This combination of SoS and HS on MRI increased cross-validated median symmetric NMI by 32% (SoS: 0·13 [0·03, 0·37] vs SoS+HS: 0·45 [0·35, 0·63], expressed as median [interquartile range]). Interpretation: We demonstrate that machine learning models using only SoS cannot unequivocally perform better than benchmarks and quantify the value of combining imaging (HS) and clinical features (SoS) in temporal epileptogenic-zone localisation. Despite good performance in localisation, no model was able to predict seizure-freedom better than benchmarks. The methods used are widely applicable, and the performance enhancements by combining clinical, imaging and neurophysiological features can be similarly quantified. Multicentre studies are required to confirm generalisability. Funding Statement: This work is supported by the Wellcome/EPSRC Centre for Interventional and Surgical Sciences (WEISS) (203145Z/16/Z). Declaration of Interests: Dr Alim-Marvasti declares no conflicts of interest. Perez-Garcia declares no conflicts of interest. Dahele declares no conflicts of interest. Dr Romagnoli declares no conflicts of interest. Dr Diehl declares no conflicts of interest. Dr Clarkson declares no conflicts of interest. Prof Duncan declares no conflicts of interest. Ethics Approval Statement: This study was approved by the Research Ethics Committee for UCL and UCLH.