Abstracts

Rationale: Predicting the occurrence of epileptic seizures using machine learning algorithms operating on non-EEG data is possible and can improve the lives of patients living with seizures. The capability to forecast seizures with minimal false alarms using data from a non-invasive wrist-worn device may permit tailoring daily activities and taking fast-acting anti-seizure medications.

Methods: We have developed a seizure forecasting algorithm trained and tested on long-term physiological data, recorded from ambulatory patients with concurrent EEG validation. The physiological data, including Accelerometer (ACC), Photoplethysmography (PPG), Electrodermal Activity (EDA), Temperature (TEMP), and Heart Rate (HR), was collected from six epileptic patients with Empatica E4 wrist band. Patients Enrolled with NeuroPace RNS Devices were given two Empatica E4 wristbands and a tablet computer to record and upload data daily to the Empatica cloud. Patients also uploaded RNS data for clinical care, and ECoG clips were reviewed for seizure activity. A Long Short-Term Memory (LSTM) Recurrent Neural Network algorithm with 4 LSTM layers, was designed and was trained on 60-second data segments from each subject. One-hour pre-ictal data segments were defined with a set-back of 15 minutes before seizure onset, and lead seizures were defined as seizures separated from preceding seizures by at least 4 hours. To compensate for the unbalanced pre-ictal/interictal data ratio in training, noise-added copies of pre-ictal data segments were generated. The data segments to train the classifier were assembled into 17 channels, including four accelerometer channels, BVP, EDA, TEMP, HR, fast Fourier transform of ACC, BVP, EDA, TEMP, HR and signal quality metrics of ACC, BVP, EDA and time of the day.

To evaluate the algorithm's performance, first, it was run individually on each human subject’s physiological E4 data, training on the initial portion of the recording and testing on subsequent data (intra-subject mode). Then the classifier was initially trained on E4 data from 55 seizures from 20 in-hospital subjects, and the last two LSTM and dense layers were retrained on training E4 data from each ambulatory subject in intra-subject mode.

Results: On average, for each patient, 8 seizures were used for training and 7 seizures for the test. The average length of the recorded data from each patient was 7 months. In intra-subject mode, training and testing on E4 data, better than chance results (p< 0.01) were obtained for 4 out of 6 patients, with an average AUC of 0.82. Using the transfer learning method and retraining the last 2 LSTM layers of the classifier, better than chance results (p< 0.05) were obtained for 4 out of 6 patients with an average AUC of 0.69.