Abstracts

Rationale: Wrist-worn devices are less intrusive than the widely used EEG systems for monitoring epileptic seizures. Data from wearable sensors can detect generalized tonic-clonic seizures. However, detection performance information regarding other seizure types is limited. Using a deep learning seizure detection model, we demonstrate detection of a broad range of seizure types by wearable signals.

Methods: Patients wore wearable sensors (E4, Empatica, Milan, Italy) at Boston Children’s Hospital’s Epilepsy Monitoring Unit between 2015 and 2017 on either wrists or ankles. We collected patients’ body temperature (TEMP), electrodermal activity (EDA), accelerometry (ACC), and blood volume pulse (BVP) data. A board-certified epileptologist determined seizure onset, offset, and seizure types using video and EEG recordings per ILAE 2017. We cleaned the data and removed seizure clusters and periods when the wearable device lost contact with skin. We applied a 10-fold patient-wise cross-validation scheme to ACC, EDA, and BVP data. Train and test sets contained different groups of patients for all model runs. We used random under-sampling and evaluated model performance on 25 clinical seizure types. We applied a convolutional neural network (CNN) based deep learning model to the raw time series sensor data to detect seizures and utilized performance measures, including sensitivity (number of seizures correctly detected using a wearable divided by the total number of seizures on EEG), False Alarm Rate (FAR; number of incorrectly detected seizures over 24 hours), and detection delay (time difference between the start of a seizure and the start of its detection).

Results: We analyzed 94 patients (57.4% female, median age 9.9 years) and 548 epileptic seizures. With a CNN based seizure detection model, all the modalities and their fusions’ performances performed better than chance (Table 1); ACC and BVP data fusion reached the best overall AUC-ROC value of 0.752, which corresponds to 18% sensitivity, 19.78 FAR and an average delay of 51.31s. Table 2 depict 25 seizure types’ AUC-ROC performance. Except in the case of Focal Non-Motor Unclassified type, all other 24 seizure types could be detected by at least 1 data modality with better than chance performance. ACC and BVP modalities reached best performance for most types of seizures.

Conclusions: Following our previous study that demonstrated that automatic epileptic seizure detection using machine learning and wearable data is feasible, we now demonstrate feasibility and better than-chance AUC-ROC performance for 24 out of 25 clinical seizure types. With the current approach, EDA overall detects seizures with longer delay and lower sensitivity than ACC and BVP. Further work will investigate the feasibility of improving detection performances through addition of clinical and physiological variables, cleaning and improving the quality of data, and utilization of larger datasets to test if the current results for seizure types with currently fewer numbers of seizures are statistically significant.

Funding: Please list any funding that was received in support of this abstract.: Epilepsy Research Fund.