Abstracts

Rationale: Seizure prediction efforts have focused on electroencephalogram data. This study aims to describe the performance of machine learning algorithms for predicting seizure duration and time of seizure occurrence based exclusively on clinical variables recorded in an electronic seizure diary. Methods: SeizureTracker.com is an online seizure diary that allows patients and families to enter basic demographic features and the duration and time of seizure occurrence. Three commonly used machine learning techniques ?"linear regression, random forests, and robust regression?" were used to train predictive models in a training subset that were then tested in an independent testing subset. Overfitting was minimized by using cross-validation in the training subset. The SeizureTracker.com dataset was divided into 65% of patients for training and 35% for testing. Algorithm performance was evaluated on predicting the 10th seizure based on the first 9, on predicting the 20th seizure based on the first 19, and on predicting the 30th seizure based on the first 29. The main outcome variables were seizure duration and times of seizure occurrence. Results: A total of 2,803 patients with a total of 63,260 seizures were used in the analyses. Median (p25-p75) age was 15 (8-30) years in the training subset and 14 (8-28) years in the testing subset. Seizure duration was one (0.3-3) minutes in both the training and testing subsets. Status epilepticus occurred in 4.5% of the training subset and 4.8% of the testing group. Interseizure interval was 1.26 (0.01-6.00) days in the training group and 0.79 (0.01-5.83) days in the testing subset. The median absolute difference between predicted and actual seizure duration was 1) for the 10th seizure was 2.26 (1.16-4.42) minutes for linear regression, 0.8 (0.15-3.9) minutes for random forests, and 0.41 (0.19-1.41) minutes for robust regression; 2) for the 20th seizure was 1.33 (0.71-3.02) minutes for linear regression, 0.62 (0.16-2.21) minutes for random forests, and 0.17 (0.07-0.88) minutes for robust regression; 3) for the 30th seizure was 1.43 (0.53-4.65) minutes for linear regression, 0.27 (0.07-1.21) minutes for random forests, and 0.39 (0.12-1.44) minutes for robust regression. The area under the curve for prediction of status epilepticus ranged from 0.69 to 0.81 depending on technique and predicted seizure (Table 1, upper portion, Figure 1A). The median absolute difference between predicted and actual time of seizure occurrence was 1) for the 10th seizure was 6.78 (5.36-8.14) days for linear regression, 9.71 (4.10-22.33) days for random forests, and 0.64 (0.48-2.47) days for robust regression; 2) for the 20th seizure was 4.98 (3.88-5.97) days for linear regression, 10.26 (4.21-25.46) days for random forests, and 0.22 (0.11-1.68) days for robust regression; 3) for the 30th seizure was 2.25 (0.93-4.61) days for linear regression, 10.27 (4.58-26.64) days for random forests, and 0.1 (0.03-1.09) days for robust regression (Table 1, lower portion, Figure 1B). Conclusions: Our data demonstrate that machine learning algorithms can predict within a reasonable range the duration and time of occurrence of seizures based on clinical data and may be of clinical utility in predicting likelihood of status epilepticus. Funding: TL and ISF are funded by the Epilepsy Research Fund.