Abstracts

Rationale: To evaluate machine learning algorithm performance for predicting time of seizure occurrence and duration with clinical variables. Methods: SeizureTracker.com is a digital seizure diary, accessible via web and mobile devices that allows patients and families to enter basic demographic features, duration and time of seizure occurrence. We used machine learning techniques to train predictive models in a training subset, subsequently tested in an independent testing subset. Performance was compared to “null” predictive models based on simple rules: for the time of seizure occurrence it was calculated evenly splitting seizures from the first to the last for each patient: the first and last seizures were assumed to occur when the first and last seizures actually occurred, but the seizure frequency was considered to have a constant interseizure interval between seizures; for seizure duration the “null” predictive model was calculated as the median of seizure duration for each patient. For seizure N duration, the predictive variables were: gender, seizure duration in seizures up to (N - 1) and time of seizure occurrence in seizures up to (N - 1). For seizure N time of occurrence, predictive variables were gender, time of occurrence in seizures up to (N - 1), and duration in seizures up to (N - 1). Time of occurrence may help patients plan their activities and duration is of interest because longer seizures are associated with poorer prognosis. All analyses were done in R (2015, version 3.2.2). Results: We analyzed the first 100 seizures in 3716 patients [2418 patients (47% males) in the training subset and 1298 patients (49% males) in the testing subset]. We excluded 6 patients (4 in the training and 2 in the testing subset) with no information on gender, 8 patients (7 in the training and 1 in the testing subset) with no information on date of birth and 32 seizures (26 in the training and 6 in the testing subset) with no information on time of occurrence. The median of the median difference between predicted and actual seizure occurrence time was best for robust regression (0.8 days), followed by MARS with no interaction (2.6 days), MARS allowing two-way interactions (2.6 days), linear regression (2.8 days), the null model (12.6 days), random forest (13.4 days), and boosted tree (27.2 days). The median of the median difference between the predicted seizure duration was best for the null model (5 seconds), followed by random forests (28 seconds), robust regression (29 seconds), boosted tree (30 seconds), MARS with no interaction (60 seconds), linear regression (102 seconds), and MARS allowing two-way interactions (1560 seconds) (Table 1).  Conclusions: Our data demonstrate that, in a large scale dataset with basic clinical data, machine learning algorithms can predict time of seizure occurrence better than a simple “null” model. The implication of our results is that time of occurrence of future seizures can be predicted better than chance using clinical data. Funding: Epilepsy Research Fund (ERF)