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

Area of Science:

• * Neurology
• * Cardiology
• * Biomedical Engineering

Background:

• * Explores the feasibility of automated seizure detection using electrocardiogram (ECG) data and machine learning.
• * Investigates the efficacy of various machine learning algorithms for identifying seizure events from ECG components.

Purpose of the Study:

• * To assess the viability of a novel automated method for seizure detection based on ECG signals.
• * To determine the optimal machine learning algorithm for maximizing seizure detection accuracy.

Main Methods:

• * Analyzed ECG recordings from 32 patients with 47 seizures using video-electroencephalogram monitoring.
• * Modeled ECG as P-Q-R-S-T heartbeats, computing derivative quantities (δX, ΔX) within sliding windows.
• * Evaluated six machine learning algorithms (Random Forest, LightGBM, XGBoost, Decision Tree, K-Nearest Neighbors, Extra Trees) using auto-ML platforms.

Main Results:

• * The Extra Trees algorithm exhibited superior seizure detection performance across validation methods.
• * A model using a 60-heartbeat window and a trigger of 20 achieved 86% sensitivity and 99.9% specificity.
• * Longer analysis windows improved sensitivity but increased detection delay.

Conclusions:

• * Automated ECG delineation is a reliable method for seizure detection.
• * High false alarm rates (1.5 per hour) necessitate further research into personalized detection algorithms.