Abstracts

Rationale: Wearable devices are less intrusive than widely used electroencephalogram systems for monitoring epileptic seizures, however their performance in detecting a variety of seizure types (ST) is limited. Using a custom-developed deep learning seizure detection model, we demonstrate detection of a broad range of ST by wearable sensors. We also present case studies on selected patients to demonstrate personalized seizure detection autoencoder models from wearable data._x000D_
Methods: Patients admitted to the epilepsy monitoring unit were enrolled and wore sensors (Empatica E4, Milan, Italy) on either wrists or ankles. We collected patients’ electrodermal activity, accelerometry (ACC), and photoplethysmography from which blood volume pulse (BVP) is derived. Two board-certified epileptologists determined seizure onset, offset, and ST using video and EEG recordings per ILAE 2017 classification.
_x000D_ We applied three generalized detection neural network models (Model 1: CNN,¹ Model 2: The CNN model trained on a larger dataset, and Model 3: CNN-LSTM) to raw time-series sensor data to detect seizures and utilized performance measures, including sensitivity, false alarm rate (FAR), and detection delay. We applied a 10-fold patient-wise cross-validation scheme to the multi-signal biosensor data and evaluated model performance on 28 ST.

We also developed a personalized autoencoder for nine patients and present three case studies to demonstrate personalized seizure detection models for each subject using wearable data. We applied a standard train-validation-test split to evaluate the personalized model._x000D_  _x000D_ Results: We included 166 patients (48% female, median age: 10 years) and 900 epileptic seizures (13254 hours of data) for 28 ST. With a CNN-LSTM-based seizure detection model, ACC, BVP, and their fusions’ performances performed better than chance (Figure 1A). ACC and BVP data fusion reached the best detection performance of 83.9% sensitivity and 35.3% FAR. Nineteen out of 28 ST could be detected by at least one data modality with AUC-ROC > 0.8 performance (Figure 1B)._x000D_ _x000D_ In case 1, the personalized models (patients 1, 2 and 7) were able to detect all generalized tonic-clonic, focal to bilateral tonic-clonic, and behavioral arrest seizures with no false alarms (FA). In case 2, we saw high detection performance for patients 4 and 6 with hyperkinetic, automatism, and focal motor tonic seizures with FAs of 12.97 and 5.14 per day respectively. In case 3, the models detected more than half of their non-convulsive seizures (patients 5, 8 and 9) with less than 100 FA (FAR < 7%) per day (Table 1)._x000D_