1275 Machine Learning for Outcome Prediction in Epilepsy Surgery: A Systematic Review

Abstract

Abstract Aim A third of epilepsy patients suffer from medically refractory seizures. In patients eligible for surgical treatment, seizure freedom rates remain variable. Machine learning (ML) utilises large datasets to detect patterns to make predictions. We systematically review studies employing ML models for prediction of outcome following resective epilepsy surgery to evaluate their efficacy, applicability and value in determining surgical candidacy. Method MEDLINE, Cochrane and EMBASE databases were searched for literature published between 2010 – 2020 according to PRISMA guidance. Non-refractory epilepsy, non-clinical outcome prediction, or non-human studies were excluded. Clinical and demographic data, ML features, discrimination and prediction accuracy metrics were extracted. Results 15 studies were included. Median cohort size was 49 (range 16 – 4211). Heterogeneous input data sources were utilised: MRI (n = 10) , electrophysiology (n = 4), PET (n = 2), clinical data (n = 2), and neuropsychological testing (n = 1). The most common ML model used was support vector machines (n = 7). All studies had good discrimination (AUC > 0.70, range: 0.79 [95% CI NR] - 0.94 [95% CI 0.92 – 0.96]), and good prediction accuracy (> 0.70, range: 0.76 [95% CI NR] – 0.95 [95% CI NR]). Limitations included small sample sizes, limited external validation and lack of comparison with clinician-predicted outcomes. Conclusions Machine Learning for outcome prediction could enhance clinical decision-making for surgical candidacy in epilepsy, and lead to improved precision medicine delivery. Outcome reporting remains inconsistent, and further work is required to externally validate such models to implement these to large-scale clinical populations.