Clinically meaningful interpretability of an AI model for ECG classification

Area of Science:

• Artificial Intelligence in Medicine
• Cardiology
• Medical Imaging Analysis

Background:

• AI models achieve high accuracy in 12-lead ECG analysis for cardiac conditions.
• Limited interpretability of AI recommendations hinders clinical adoption of these tools.
• A need exists for interpretable AI solutions in digital ECG diagnostics.

Purpose of the Study:

• To demonstrate the feasibility of a generic interpretability tool for AI models analyzing digitized 12-lead ECG images.
• To provide medically relevant interpretations of AI classifications for physicians.
• To enhance the clinical integration of AI in ECG analysis.

Main Methods:

• Utilized Jacobian matrix sensitivity to compute classifier gradients for pixel-wise interpretability.
• Developed a generic tool applicable to AI models analyzing 12-lead ECG images.
• Validated methodology on a large dataset of scanned and mobile-captured ECG images.

Main Results:

• The interpretability tool highlighted diagnostically relevant ECG features for morphological and arrhythmogenic conditions.
• The tool identified significant signal features indicating the absence of specific cardiac conditions.
• Achieved high correlation between the AI interpretability method and expert electrophysiologist interpretations.

Conclusions:

• The developed tool offers feasible, medically relevant interpretability for AI-based ECG analysis.
• This approach can improve physician understanding and trust in AI diagnostic recommendations.
• Enhances the potential for seamless clinical integration of AI in cardiology practice.