Optimizing beat-wise input for arrhythmia detection using 1-D convolutional neural networks: A real-world ECG study.

Area of Science:

• Cardiology
• Biomedical Engineering
• Machine Learning

Background:

• Cardiac arrhythmias present diagnostic challenges in real-world settings.
• Machine learning aids arrhythmia detection, but optimal input size is understudied.
• Evaluating performance across inter-patient and intra-patient conditions is crucial.

Purpose of the Study:

• Determine the optimal number of heartbeats for accurate arrhythmia classification.
• Assess machine learning model performance under varying patient conditions.
• Evaluate performance-resource trade-offs for mobile health (mHealth) deployment.

Main Methods:

• Preprocessed electrocardiography (ECG) signals using beat-wise segmentation and resampling.
• Employed a 1-D convolutional neural network for classifying eight multi-labeled arrhythmias.
• Validated model performance using fivefold cross-validation on combined real-world and database ECG data.

Main Results:

• Peak accuracy (94.82%) achieved with four heartbeats under inter-patient conditions.
• Minimal performance gains observed with more than four beats.
• Demonstrated feasibility for real-time detection in large-scale simulations (approx. 5000 patients).

Conclusions:

• Four heartbeats offer an optimal balance between accuracy and computational efficiency for arrhythmia classification.
• Findings are critical for real-time wearable ECG devices with resource constraints.
• Provides a reference for designing scalable and efficient mHealth arrhythmia detection systems.