Transparent and Robust Artificial Intelligence-Driven Electrocardiogram Model for Left Ventricular Systolic...

Area of Science:

• Cardiology
• Artificial Intelligence in Medicine
• Medical Device Software

Background:

• Heart failure (HF) poses a global health challenge, with early detection hindered by traditional diagnostic limitations.
• Deep learning offers novel approaches for identifying left ventricular systolic dysfunction (LVSD), a critical HF indicator, using electrocardiogram (ECG) data.

Purpose of the Study:

• To validate AiTiALVSD, an AI-enabled ECG Software as a Medical Device, for accuracy, transparency, and robustness in detecting LVSD.
• To assess the clinical plausibility and reliability of AI-driven LVSD detection.

Main Methods:

• Retrospective single-center cohort study of 688 patients suspected of LVSD.
• Evaluation of the AiTiALVSD deep learning model against echocardiographic ejection fraction.
• Application of Testing with Concept Activation Vectors (TCAV), clustering, and robustness testing for transparency and reliability.

Main Results:

• AiTiALVSD achieved high diagnostic performance with an AUROC of 0.919.
• Significant correlation observed between AiTiALVSD scores and left ventricular ejection fraction.
• TCAV analysis confirmed model alignment with medical knowledge; robustness to ECG noise showed a specificity decrease.

Conclusions:

• AiTiALVSD demonstrates high diagnostic accuracy, transparency, and resilience for early LVSD detection.
• The AI tool shows potential for clinical application in early heart failure diagnosis.
• This study emphasizes the importance of transparency and robustness in AI-ECG for advancing cardiac care.