Explainable ECG Analysis by Explicit Information Disentanglement With VAEs.

Area of Science:

• Artificial Intelligence
• Cardiology
• Machine Learning

Background:

• Electrocardiogram (ECG) interpretation is crucial for diagnosing cardiac conditions but traditionally relies on time-consuming expert analysis.
• Existing AI models for ECG analysis often lack the explainability required for clinical application.
• Subtle features in ECG signals can be missed by traditional methods, impacting diagnostic accuracy.

Purpose of the Study:

• To develop an explainable AI (XAI) method for ECG classification and interpretation.
• To enhance the clinical utility of AI in cardiology by providing model explainability.
• To predict Left Ventricular Function (LVF) from ECG signals using a novel XAI approach.

Main Methods:

• A variational autoencoder (VAE) latent space was partitioned into label-specific and non-label-specific subsets.
• An adversarial network constrained one subset from learning label-specific information, enabling supervised disentanglement.
• The disentangled latent space was used for attribute manipulation to visualize ECG features and predict LVF.

Main Results:

• The proposed XAI method effectively segregated LVF-specific information in the VAE latent space.
• The model achieved improved prediction performance over state-of-the-art VAE methods (AUC 0.832 vs. 0.790, F1 0.688 vs. 0.640).
• The model demonstrated comparable performance to ground truth LVF in predicting survival (concordance 0.72).

Conclusions:

• The developed XAI model facilitates LVF prediction interpretation by providing visual context to ECG signals.
• This approach offers a generalizable method for explainable and predictive AI in ECG analysis.
• The explainable AI model has the potential to reduce the time and expertise needed for ECG analysis, aiding clinical decision-making.