Abstracts

Rationale: A major source of anxiety and frustration for people with epilepsy is the uncertainty in knowing when the next seizure may occur. There are currently no methods available for patients to have a meaningful seizure risk forecast for the coming 24-hours. Methods: Custom software using Python 2.7.14, Keras 2.1.5, and Tensorflow 1.7.0 was developed. Data was exported from SeizureTracker.com, one of the world’s largest patient-reported seizure diary databases, from December 1, 2007 through November 30, 2015. After excluding patients reporting seizure dates outside that range as well as negative seizure durations and patients with th of the seizure period. Each training input (except those with zeros seizures) were convolved with the unique kernel to expand the forecast “umbrella”. These inputs were used as the training data for a deep learning architecture (40,609 free parameters) that utilized recurrent neural networks, allowing the model to be repeatedly applied to the entire diary of each patient, making a single 24-hour prediction at a time with a moving window (using 83 days of history for any given forecast). The loss function was mean squared error, batch size was 10,000 and 5 epochs were run. The resulting model was tested on the test set (which did not have any convolution applied). Results: 1726 patients met inclusion/exclusion criteria. 930347 predicted days were used for the training set, and 335286 days were used in the test set. The entire test set included 77.2% days with no seizures. The forecasts suggested 60.3% days with no seizures. Thresholds were set to 0.1 for NO SEIZURE and 0.99 for SEIZURE. Across patients, the median specificity was 98.1%, sensitivity 32.3%, and accuracy 89.3%. Accounting for the held out patients only (i.e. those that were entirely excluded from the training set), medians were 92.0% specificity, 65.2% sensitivity, and 85.2% accuracy. A random predictor of no-seizure days with a matched frequency for each patient had a median specificity of 81.3%. Conclusions: It may be possible to forecast the 24-hour future seizure risk using a 83-day seizure history alone. Forecasting “no seizure” days can be done with high specificity. Accurate prediction of seizure count for days that do include seizures was not achieved, though future modifications to the algorithm may allow for this. These forecasts may have clinical utility alone, or they could be used as priors for biosensor-based forecasting systems. The overall utility of this method requires further validation and study. Funding: This work was supported in part by NIH T32NS048005, Neurostatistics and Neuroepidemiology Fellowship.