DeepTWA-TM: Deep Learning T-Wave Alternans Detection in Ambulatory ECG via Time Analysis.

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Sudden cardiac death risk assessment relies on non-invasive markers like T-wave alternans (TWA) from ECG.
• Clinical TWA detection is challenging due to standardization issues and variable ambulatory ECG recording conditions (noise, artifacts).

Purpose of the Study:

• To develop a Deep Learning (DL) approach for direct T-wave alternans (TWA) detection from electrocardiogram (ECG) signals.
• To simplify the TWA detection pipeline by removing the need for traditional signal processing steps.
• To enhance the practical applicability of TWA analysis for sudden cardiac death risk stratification.

Main Methods:

• Utilized transfer learning with established DL architectures (VGG, ResNet, Inception) for TWA detection.
• Trained models on a custom, long-term dataset of real patient ECGs, including micro-alternans and higher amplitude TWA (20-100 μV).
• Employed patient separation during training and testing to ensure model generalizability and robust performance.

Main Results:

• The DL model achieved a significant F1-score of 0.83 in ambulatory TWA analysis.
• The proposed method outperformed traditional Machine Learning approaches in TWA detection accuracy.
• Eliminated the need for R-peak identification, T-wave segmentation, and feature engineering, simplifying the process.

Conclusions:

• The DL approach offers a simplified and adaptable method for TWA detection directly from ECG signals.
• This advancement facilitates more efficient and effective risk stratification for sudden cardiac death in clinical settings.
• The model's performance in ambulatory conditions demonstrates its potential for real-world application.