Abstracts

Automated Detection of Convulsive Seizures Using Video Recordings with Privacy Preserving Features

Abstract number : 2.001
Submission category : 3. Neurophysiology / 3A. Video EEG Epilepsy-Monitoring
Year : 2023
Submission ID : 383
Source : www.aesnet.org
Presentation date : 12/3/2023 12:00:00 AM
Published date :

Authors :
Presenting Author: Shobi Sivathamboo, PhD – Monash University

Hugh Simpson, MBBS, PhD – Monash University; Deval Mehta, PhD – Monash University; Theekshana Dissanayake, BSc – Monash University; Lyn Millist, BAppSc – Alfred Health; Zongyuan Ge, PhD – Monash University; Terence O'Brien, MD – Monash University; Patrick Kwan, MD, PhD – Monash University

Rationale: Automated seizure detection devices for out-of-hospital epilepsy care are needed. However, current single-purpose and dedicated wearable devices have limitations. Here, we investigated the performance of automated video-based seizure detection approaches that preserve patient privacy.

Methods: We analyzed convulsive seizures recorded from patients with epilepsy during inpatient video-EEG monitoring admission who had with video and electrocardiogram (ECG) data. Four distinct epochs were identified for each seizure including baseline (60s, ending 240s before seizure onset), pre-ictal (60s, ending at seizure onset), ictal (seizure onset to termination), and post-ictal (300s, starting from seizure termination). We extracted the privacy preserving feature of optical flow, which is a measure of motion, from each video recording. We used two different transformer deep neural networks to analyze features of the video and raw ECG data. Our findings were validated using a five-fold, patient-independent, cross-validation approach.

Results: Forty seizures (35 focal-to-bilateral and 5 generalized tonic-clonic seizures) from 31 patients were included. Using video data alone, the model demonstrated a mean sensitivity of 85% and specificity of 87%; while ECG alone yielded a sensitivity of 97% and specificity of 88% for seizure compared to respective baseline epochs (Figure 1). When video and ECG data were combined, the performance of the model improved to a sensitivity of 95% and specificity of 95%.

Conclusions: We demonstrate high sensitivity and specificity for the automated detection of convulsive seizures using video and ECG. Importantly, the video data was extracted and processed using privacy-preserving methodology. These data can be derived from existing commercially available technology, making them more cost-effective, non-stigmatizing to users, and accessible to most patients with epilepsy. Future studies should compare the performance of video-based methods for seizure detection against current non-invasive wearable devices.

Funding: Australian National Health and Medical Research Council (NHMRC).



Neurophysiology