Abstracts

Rationale: It is well known that seizure occurrences can be associated with external factors, such as the time of day. However, most attempts to forecast or predict seizures rely on features derived only from EEG signals, especially intracranial EEG. This study explores the use of non-EEG features to forecast seizures: time of day and weather, as well as rate of spike-wave discharge, which is easily measured using scalp EEG. using scalp EEG. Methods: Data were from nine patients with focal epilepsy whose intracranial EEG had been continuously recorded for 2.1 years on average (Cook, M.J. et al., 2013, Lancet Neurol, 12(6), pp. 563-71). Seizure likelihood distributions were calculated for each patient from a training set based on time of day, air temperature, humidity, wind speed, rainfall, and spike rate. Seizure likelihood from each of these factors was then determined for test samples. Likelihoods were also combined using a Bayesian approach to produce a single forecast per sample, where each sample consisted of a 10 minute intervention/warning time and a 10 minute seizure occurrence period. Forecasts are predictions of the likelihood that a seizure will occur within the seizure occurrence period of a given sample (range 0-1). Performance was measured by area under the curve (AUC) of the receiver-operator characteristic (range 0 – 1). Generally, an AUC above 0.95 indicates forecasting performance good enough to be useful to the patient, although this value will change between patients depending on their personal needs and seizure frequency. Results: Patients 8, 10, 11, and 15 showed the best results when combining all available factors (AUC = 0.886, 0.761, 0.729, and 0.806, respectively). Results for patients 3, 6, and 13 were the best when combining spike information with weather (AUC = 0.953, 0.906, and 0.714 respectively), and for Patients 1 and 9 were best with time of day alone (AUC = 0.520 and 0.912, respectively). Conclusions: For most patients, forecasting was better when multiple external factors were combined compared to any single factor. However, performance was not yet good enough to be useful for most patients. With a more extensive selection of features to draw from, the techniques used here may improve seizure forecasting to a level that provides real benefit to patients. Funding: National Health and Medical Research Council, Project Grant ID No: 1065638 University of Melbourne, Melbourne Research Scholarship