Detecting cardiovascular diseases using ECG scans and explainable artificial intelligence.

Area of Science:

• Artificial Intelligence in Healthcare
• Cardiovascular Disease Diagnostics
• Medical Image Analysis

Background:

• Cardiovascular diseases are a leading global cause of death.
• Electrocardiography (ECG) is crucial for early detection.
• Current AI models for ECG analysis lack transparency and robustness to real-world image variations.

Purpose of the Study:

• To introduce an explainable AI framework for quantifying the stability of deep learning models used in ECG analysis.
• To identify vulnerabilities in AI models when subjected to controlled image perturbations.
• To enhance the reliability and clinical applicability of AI in cardiovascular diagnostics.

Main Methods:

• Utilized a large-scale dataset of synthesized ECG printouts (PTB-XL benchmark) with clean and manipulated versions.
• Trained and evaluated four deep learning architectures (EfficientNet, InceptionNet) with varying activation functions.
• Assessed model stability using local interpretable model-agnostic explanations and intersection over union metrics.

Main Results:

• Models trained on augmented datasets showed improved generalization to perturbed data (AUC 0.894).
• Stability analysis revealed higher explanation consistency (IoU 0.399) for models trained on perturbed data.
• Radiomic-like features accurately identified deep learning models (up to 98% accuracy).

Conclusions:

• The developed framework provides visual and quantitative evaluation of AI stability in ECG analysis.
• Identifies how image manipulations impact AI reliability, guiding development of robust algorithms.
• Promotes reproducible research and collaboration through publicly available tools and datasets.