Assistant Professor University of North Florida, Florida
Rationale: The rapid growth of wirelessly connected physiological monitors is expanding possibilities in ambulatory EEG monitoring and responsive neuromodulation in epilepsy. With increases in mobile computational power comes the ability to run more complex algorithms for seizure diaries and closed-loop neuromodulation using a broader range of physiological data. Application of trained algorithms across different subjects would be very useful, but prior studies suggest cross-subject application is less accurate than intra-subject training and testing. Transfer learning has recently emerged as a valuable technique for leveraging large repositories of available data to improve performance and reduce training time in scarce data sets, though few reports exist evaluating these methods for time-series data. Methods: We have developed a generalized seizure detection algorithm trained with in-hospital and ambulatory iEEG signals, re-trained using wearable device data from subjects, and tested across different patients with epilepsy. A Long Short-Term Memory (LSTM) Recurrent Neural Network (RNN) algorithm was designed and was pre-trained on ten-second iEEG segments from 38 human subjects undergoing pre-surgical invasive in-hospital EEG monitoring. The algorithm was re-trained and cross-validated on ambulatory NeuroVista (NV) iEEG data from 10 human and canine subjects with epilepsy. To compensate for the unbalanced ictal/interictal data ratio in the iEEG recordings, noise-added copies of ictal data segments were generated for training. The initial algorithm, pre-trained on 38 patients, was then adaptively re-trained to detect focal motor and generalized tonic-clonic seizures in wrist-worn data from ten patients with epilepsy. The wrist-worn Empatica E4 records 3-axis accelerometery, photoplethysmography (PPG), electrodermal activity (EDA), temperature (TEMP) and heart rate (HR). Empatica E4 data was collected from in-hospital patients undergoing pre-surgical scalp or intracranial EEG monitoring. We evaluated the performance of the algorithm at each stage of training. To provide a baseline for comparison, the algorithm was first run in an intra-subject fashion on each canine and human subject’s NV iEEG data, training on the initial portion of the recording and testing on subsequent data. The same LSTM network architecture was trained on epileptologist- marked seizures and interictal epochs in iEEG data from 38 ICU patients, and was then adaptively re-trained on 16-channel ambulatory data from nine NV subjects and tested on each subject in a patient-wise cross validation, and performance metrics were calculated. The classifier was then adaptively re-trained on Empatica E4 data, including the FFT of accelerometer, PPG and EDA channels on three groups of seven subjects, and was tested on the three remaining subjects. In all experiments algorithm performance was evaluated using the area under the ROC curve (AUC) and area under the precision-recall curve (AUPR). Results: The mean AUC for conventional intra-subject training and testing for the NV subjects was 0.96. The mean AUC for the cross-patient experiment with initial training on the ICU data and adaptive retraining for the NV subjects was 0.88 (0.72 per-segment). The average AUC of the cross patient test, with re-trained classifier, on the Empatica E4 data from 12 patients was 0.97. Conclusions: Although cross-subject application is less accurate than intra-subject training and testing, transfer learning is a valuable technique for using available data to improve performance and reduce training time. Funding: Please list any funding that was received in support of this abstract.: This work was funded by the ‘My Seizure Gauge’ grant provided by the Epilepsy Innovation Institute, a research program of the Epilepsy Foundation of America.