Deep neural networks learn by using human-selected electrocardiogram features and novel features

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Signal Processing

Background:

• Deep neural networks (NNs) offer advanced capabilities in electrocardiogram (ECG) analysis.
• Understanding the features learned by AI models is crucial for clinical trust and interpretability.
• Human experts utilize specific features for ECG interpretation, providing a benchmark for AI performance.

Purpose of the Study:

• To investigate if AI, specifically NNs, for ECG analysis can be explained using human-selected features.
• To quantify the explainability of AI models in ECG interpretation.
• To determine if AI models learn features comparable to those of human experts.

Main Methods:

• Utilized a dataset of 100,000 ECGs annotated with human-explainable features.
• Applied linear and non-linear models to predict AI model outputs for age and sex detection.
• Employed canonical correlation analysis to quantify shared information between NN and human features.
• Reconstructed human-selected ECG features from AI-derived features using linear models.

Main Results:

• Strong correlations (0.49-0.70) were observed between simple models and AI outputs.
• Human explainable features showed high correlation (>0.85) with key AI-identified age and sex features.
• Single human-selected ECG features were linearly reconstructed from AI features with high accuracy (up to 0.86).

Conclusions:

• NNs for ECG analysis extract features similarly to human experts.
• AI models generate novel features beyond human expertise, leading to superior performance.
• The study validates the explainability of AI in ECG analysis and its alignment with human expert feature extraction.