Abstract Details

Title Novel Method to Classify Epileptic and Non-Epileptic Seizures Using Wearable Watch Data

Topic Epilepsy/Clinical Neurophysiology (EEG)

Presentation(s) P5 - Poster Session 5 (8:00 AM-9:00 AM)

Poster/Presentation
Number 9-007

Objective To demonstrate the feasibility of using wearable data to distinguish between epileptic and non-epileptic seizures using a modified Convolutional Neural Network (CNN) model.

Background Distinguishing psychogenic non-epileptic seizures (PNES) from epileptic seizures (ES) is a major challenge. Nearly 30% of those referred to epilepsy centers experience non-epileptic events. Misdiagnosis can lead to inappropriate treatment and prolonged hospitalization. The gold standard for diagnosis is video recordings and EEG, but these methods are expensive and cumbersome. Wearables are less invasive and more portable, thus offering a promising alternative for seizure diagnosis. Convolutional Neural Networks (CNNs) have shown potential potential in analyzing EEG signals for seizure classification and have been used to analyze electrodermal activity accurately. However, there is a need for studies that use CNNs on data from wearables only.

Design/Methods Data was collected from patients (n=29) at Baylor St. Luke’s Epilepsy Monitoring Unit, who wore Empatica E4 wristbands. Collected data included skin temperature (TEMP), electrodermal activity (EDA) and heart rate (HR). Data segments from immediately after seizure events to 60 minutes after were used. The study design involved training a modified 1-dimensional ResNet-18 model in two stages: training and validation on a modified subset of data that had a 1 to 1 ratio of ES vs. PNES samples, then training and validation in a leave-one-patient-out manner. Lastly, the model was tested on a test set that had an ES:PNES ratio reflective of the whole dataset.

Results The model demonstrated high accuracy and AUC-ROC. Test set metrics included an AUC-ROC of 76.9%, accuracy of 75.6%, precision of 64.3%, recall of 69.1%, and F1 Score of 65.5%.

Conclusions The CNN model's performance in distinguishing between epileptic and non-epileptic seizures using wearable data shows promising potential. With continued refinement and increased data, this method could significantly enhance seizure monitoring in clinical and outpatient settings.

Authors/Disclosures
Jennifer Zhou, BS PRESENTER	Miss Zhou has nothing to disclose.
Peikun Guo	Mr. Guo has nothing to disclose.
Sruthi Gopinath Karicheri, Programmer/Analyst	Mrs. Gopinath Karicheri has nothing to disclose.
Akane Sano, PhD	The institution of Dr. Sano has received research support from National Science Foundation. The institution of Dr. Sano has received research support from National Institute of Health.
Zulfi Haneef, MD, MBBS, MRCP, FAAN	Dr. Haneef has received publishing royalties from a publication relating to health care.

��ɫ��

��ɫ��