Different Ventricular Fibrillation Types in Low-Dimensional Latent Spaces.

Area of Science:

• Cardiology
• Computational Biology
• Machine Learning

Background:

• Ventricular fibrillation (VF) mechanisms remain unclear, with conventional analysis lacking discriminative features.
• Identifying distinct VF patterns is crucial for understanding its underlying causes.

Purpose of the Study:

• To investigate if low-dimensional latent spaces can reveal discriminative features for different VF mechanisms.
• To assess the utility of manifold learning with autoencoder neural networks for VF analysis.

Main Methods:

• Surface ECG recordings from an animal model during VF episodes (onset to 6 min).
• Analysis of manifold learning using autoencoder neural networks.
• Inclusion of control, drug interventions (amiodarone, diltiazem, flecainide), and autonomic blockade conditions.

Main Results:

• Latent spaces from unsupervised and supervised learning showed moderate separability for different VF types.
• Unsupervised schemes achieved 66% multi-class classification accuracy.
• Supervised schemes improved separability, reaching up to 74% classification accuracy.

Conclusions:

• Manifold learning in low-dimensional latent spaces provides valuable, separable features for studying VF types.
• Machine-learning-generated latent variables are superior VF descriptors compared to conventional time or frequency domain features.
• This technique aids in elucidating underlying VF mechanisms.