Improving explainability of deep neural network-based electrocardiogram interpretation using variational...

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Informatics

Background:

• Deep neural networks (DNNs) show promise in interpreting electrocardiograms (ECGs) for various applications, including ejection fraction (EF) assessment.
• Clinical implementation of DNNs is hindered by a lack of explainable techniques, with current heatmap methods proving unreliable.

Purpose of the Study:

• To develop and evaluate a novel, explainable AI pipeline for ECG interpretation that overcomes the limitations of current 'black box' models.
• To enhance clinician trust and facilitate the adoption of AI in cardiovascular diagnostics through transparent algorithmic reasoning.

Main Methods:

• A variational auto-encoder (VAE) was employed to learn interpretable ECG morphology factors (FactorECG).
• These factors were integrated into common prediction models, enabling explainable ECG analysis at both individual and model levels.
• The pipeline was trained on a large dataset of 1.1 million ECGs, compressing them into 21 physiologically relevant factors.

Main Results:

• The explainable FactorECG pipeline demonstrated performance comparable to 'black box' DNNs across conventional ECG interpretation (AUROC 0.94 vs. 0.96), reduced EF detection (AUROC 0.90 vs. 0.91), and 1-year mortality prediction (AUROC 0.76 vs. 0.75).
• Unlike DNNs, the FactorECG pipeline provided clear insights into the specific ECG morphological changes driving predictions.
• These findings were validated in an external, population-based dataset.

Conclusions:

• Explainable AI pipelines are crucial for the clinical implementation of DNNs in ECG analysis.
• Adopting transparent AI methods will foster greater clinician confidence and enable the identification of potential model biases.
• Future research should prioritize the development and integration of explainable AI in cardiovascular diagnostics.