Abstracts

Rationale: Persons with epilepsy (PWE) have a 3- to 4-fold risk of death as compared to the general population, and the risk is greatest during the first two years after developing epilepsy. The overall risk of death appears to be multifactorial, and many potential predictors of mortality have not been adequately explored. Analyses of large clinical datasets (big data) may disclose clinical predictors of a person’s risk of death that cannot be gleaned from typical datasets derived from smaller cohort studies. We hypothesized that the individual risk of death in PWE may be predicted using machine learning (ML) algorithms and large, population-based health data.Methods: PWE were identified in The Health Improvement Network (THIN) database from the UK using a validated case definition (i.e. at least one code for epilepsy syndromes or two codes for epilepsy symptoms, and a subsequent prescription for an antiepileptic drug (AED)) (n=34,465). Patient features (covariates) that were hypothesized to be predictive of mortality based on a literature search and consultation with clinicians were then extracted. For each patient, we extracted 86 features, including: types of seizures, number of AEDs tried and age at diagnosis. These feature vectors were used to train a linear support vector machine to predict mortality. We created predictive models for different time periods (0-2 years, 10-17 years and 0-20 years after diagnosis) to see if the model performance varied based on survival time. For the applicable time periods, we trained and tested the model using 5-fold stratified cross-validation with patient information that was collected at different time points (at diagnosis, 1 year and 5 years after diagnosis) to determine if the model’s predictive success could be increased by using information that would be available at later time points. The predictive performance of each model was derived using sensitivity, specificity, ROC, F1 score, and negative and positive predictive values.Results: Models were predictive of mortality with varying accuracy. The model predicting mortality in the first two years using the information available at diagnosis had an F1 score of 0.777. The F1 scores for the models at 10-17 years after diagnosis were between 0.716 and 0.729; while those for the 20-year period from ranged from 0.924-0.974. Using data from long observation periods after diagnosis did not significantly change the model’s accuracy. Age at diagnosis, high antiepileptic drug (AED) daily dose, poor adherence to AEDs (calculated), and having convulsions or status epilepticus were most predictive of mortality.Conclusions: The results suggest that it is possible to develop a tool that allows clinicians, at the point of clinical encounters, to assess risk of mortality in individual PWE, and to intervene more effectively by taking precautionary steps to mitigate this risk, and by counselling those at risk and their families. Future research will explore different ML algorithms and predictors, and their performance in alternative datasets.