Abstracts

Rationale: The objective was to develop a flexible software system, for use in epilepsy diagnostic and treatment centers that require highly reliable short-term prediction and/or very early detection of epileptic seizures. We created the word “predetection” to have this meaning.Methods: Advanced signal processing and recurrent neural net (RNN) software extracts various parameters from multi-channel scalp EEG recording to predetect seizures. The system utilizes automated feature selection, with the selected features being fed to a RNN. Feature selection, channel selection and RNN training is patient specific, using a sparse set of expert-defined ictal, preictal and interictal periods. Results: We analyzed 795 hours of data from 25 different patients, exhibiting a variety of seizure manifestations. The seizure predetection software was tested on long continuous recordings ranging from 15 to 62 hours (average 32 hours). The EEG data used were recorded from epilepsy patients hospitalized for long-term EEG monitoring as part of pre-surgical evaluation of seizure pattern and localization in five cooperating centers around the country. All of the 25 patients had localization-related epilepsy and most ranged in age from 20 to 40 years (outliers at 5 and 58 yr). Each patient had 2-5 seizures, one of which was used for training (in five cases two were used). Testing was done with 46 seizures, which included none that were used for training. All 46 seizures were complex partial (CPS) onset of which 12 generalized (CPS+G). The sensitivity of the seizure warning system was 100% (all seizures successfully detected).The median warning time from detection to electrographic seizure onset was 32 sec for the CPS group and -1 sec for CPS+G group. This is consistent with warning time being a function of propagation time in localization-related epilepsy. False alarm rate was no more than 2 per day (0.083/hour) in 80% of patients. Effective training was typically performed on one seizure. Of 25 patients, 18 had positive predictive value (PPV) of 0.5 or more and 12 had a perfect PPV of 1.0, indicating 100% accuracy. Conclusions: These findings suggest that this seizure predetection system is capable of early seizure onset detection with performance characteristics that could have practical clinical utility. (Supported by NIH/NINDS SBIR Grants R44NS039214 and R43NS051881.)