Authors :
Presenting Author: Deval Mehta, PhD – Monash University
Shobi Sivathamboo, PhD – Monash University
Talha Ilyas, MS – Monash University
Ilma Wijaya, BS – Monash University
Rob Steele, BS – Monash University
Alison Conquest, BBiomedSc (Hons.) – Monash University
Lyn Millist, BS – Alfred Health
Hugh Simpson, PhD – Monash University
Theekshana Dissanayke, PhD – Technical University of Berlin
Jarrel Seah, MD – Alfred Health
Zongyuan Ge, PhD – Monash University
Terence J. O'Brien, MBBS, MD – Alfred Health
Patrick Kwan, MD PhD – Monash University
Rationale:
Functional seizures are often misdiagnosed as epileptic seizures because of their similar behavior, leading to unnecessary treatment with antiseizure medications and delaying appropriate intervention. Machine learning models based on video and electrocardiogram (ECG) data may help distinguish functional from epileptic seizures in the community. However, analyzing raw videos poses privacy risk for the individuals. This project aimed to develop automated video and ECG-based seizure detection models using a privacy preserving approach.
Methods:
We selected video and ECG recordings of tonic-clonic seizures (TCS) and convulsive functional seizures from patients admitted to the epilepsy monitoring unit. Four distinct epochs were extracted for analysis for each seizure, including baseline (60s, ending 240s before seizure onset), pre-ictal (60s, ending at seizure onset), ictal (behavior seizure onset to termination), and post-ictal (300s, starting from seizure termination). We extracted the privacy preserving features of 17 body joints (including torso, arms, legs), 21 hand joints, 70 facial keypoints, and optical flow (a measure of motion) from each video recording. We trained and internally validated separate transformer deep learning models that combined the extracted video features and ECG data for the identification of TCS and functional seizures and normal activity, and for differentiating between TCS and functional seizures.
Results:
38 TCS (33 focal to bilateral tonic-clonic seizures and 5 generalized tonic-clonic seizures) from 31 patients with epilepsy and 40 functional seizures from 40 individuals without epilepsy were included. The median duration of TCS was 112 seconds (range 55s to 513s) and of functional seizures was 230 seconds (range 25s to 961s). Our models (Figure 1): (a) identified 97.4% (37/38) of TCS and 94.7% (36/38) of normal activity recordings from the patients with epilepsy; and (b) identified 92.5% (37/40) of functional seizures and 85.0% (34/40) of normal activity recordings from the individuals with functional seizures; and (c) distinguished 97.5% (39/40) of functional seizures and 94.7% (36/38) of convulsive seizures.Conclusions:
We have developed machine learning models to accurately identify and differentiate between TCS and convulsive functional seizures using privacy preserving video features and ECG data. Future research should evaluate the false alarm rate of the models in continuous recordings and validate the models on external datasets including home recordings. By incorporating these models into commercial video and ECG recording devices, they may help shorten the time to correct diagnosis of functional seizures and enable earlier treatment, leading to improved patient outcomes.
Funding:
Australian National Health and Medical Research Council (NHMRC).