Electrocardiogram (ECG)-based seizure detection using supervised machine-learning

We conducted a pilot study utilizing automatic delineation of electrocardiogram (ECG) and machine learning that considered all components of the ECG complex for seizure detection. The primary outcome was to assess the feasibility of this method. The secondary outcome was to identify the most effective machine learning algorithm.

We screened ECG recordings from patients included in the EPICARD cohort who underwent video-electroencephalogram monitoring. A total of 47 seizures from 32 patients were selected. Epochs of 90 min surrounding the seizures were retained. Each ECG was converted into a sequence of heartbeats modeled as a P-Q-R-S-T succession. Derivative quantities measuring time variations between the inner and outer components of heartbeats were computed, designated as δ_X and ΔX. Our algorithm monitored 3 to 60 successive heartbeats within a sliding window. An alarm was triggered when more than N heartbeats were classified as in-seizure (N between 3 and 20). Heartbeats were categorized as in-seizure by trained neurophysiologists. We used automated machine learning (auto-ML) platforms (Dataiku & Flaml) to assess six different algorithms: Random Forest, LightGBM, XGBoost, Decision Tree, K-Nearest Neighbors, and Extra Trees.

The Extra Trees algorithm provided the best seizure detection performance regardless of the validation method used. Although longer-window models enhance detection sensitivity, they do so at the cost of delayed identification. A model analyzing 60 heartbeats with a trigger of 20 achieved 86 % sensitivity and 99.9 % specificity.

Automatic delineation is reliable, however the false alarm rate remains high (1.5 per hour). Future work should focus on personalizing detection algorithms to improve this false alarm rate.