Interpretable manifold learning for T-wave alternans assessment with electrocardiographic imaging

Area of Science:

• Cardiology
• Biomedical Engineering
• Data Science

Background:

• T-wave alternans (TWA) is a key biomarker for predicting sudden cardiac death.
• Electrocardiographic imaging (ECGI) provides high-resolution epicardial data for TWA analysis.
• Current TWA methods often ignore spatial information from ECGI, analyzing signals independently.

Purpose of the Study:

• To develop a novel, subject-specific, and interpretable TWA estimation method utilizing ECGI spatial data.
• To improve the accuracy and personalization of TWA analysis for sudden cardiac death risk stratification.

Main Methods:

• Implemented a manifold learning approach using Uniform Manifold Approximation and Projection (UMAP) for dimensionality reduction.
• Employed the Louvain algorithm to identify TWA-dominant communities within the ECGI data.
• Utilized Bootstrap analysis for classification and Shapley additive explanations for interpretability.

Main Results:

• Dimensionality reduction to 18 dimensions improved TWA-dominant community separation (average normalized distance of 0.28).
• Bootstrap analysis indicated TWA-dominant communities exceeded confidence intervals.
• Identified distinct TWA patterns (hump-shaped, amplitude-shifted) and confirmed UMAP's focus on these features.

Conclusions:

• This is the first TWA detection method specifically designed for ECGI data.
• The subject-specific approach allows for personalized diagnostic insights by extracting individual characteristics.
• The method demonstrates high accuracy and consistent patient decision-making, enhancing sudden cardiac death risk assessment.