MnL-TWA: Manifold Learning Approach for T-Wave Alternans Detection in Ambulatory Environments.

Area of Science:

• Cardiology
• Machine Learning
• Data Science

Background:

• T-wave alternans (TWA) signifies ECG variations linked to cardiac instability and sudden cardiac death risk.
• Current machine learning (ML) methods for TWA detection lack transparency due to their black-box nature.

Purpose of the Study:

• To improve the explainability of ML models for TWA detection using manifold learning (MnL).
• To maintain or enhance TWA detection effectiveness while increasing model interpretability.

Main Methods:

• Fine-tuning nonlinear dimension reduction techniques (UMAP, Isomap, AE) with ML classifiers (KNN, RF, NN).
• Evaluating performance using patient-wise permutations for robust assessment.
• Utilizing latent space visualization to understand classification decisions.

Main Results:

• Autoencoder-based neural networks (AE-NN) retained key discriminative information (F1-score 92.1 ± 2.4 %).
• MnL spaces revealed misclassifications near decision boundaries, linked to low-amplitude, dispersed TWA voltages.
• Isomap-RF and AE-NN achieved TWA detection performance comparable to traditional methods (F1-score ~78%).

Conclusions:

• MnL-generated spaces offer insights into TWA classification and amplitude patterns.
• This approach bridges the gap between ML model performance and transparency.
• MnL supports more clinically reliable TWA detection by enhancing interpretability.