Non-Invasive Diagnosis of Chronic Myocardial Infarction via Composite In-Silico-Human Data Learning.

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Myocardial infarction (MI) is a major global cause of death.
• Accurate infarct quantification is vital for patient management.
• Current gold standard, LGE-CMR, uses toxic gadolinium contrast agents, posing risks for patients with kidney disease.

Purpose of the Study:

• To develop a completely non-invasive method for identifying infarct location and extent in the left ventricle.
• To utilize machine learning (ML) with cardiac strain data as input.
• To overcome limitations of invasive contrast agents in cardiac imaging.

Main Methods:

• A multi-fidelity ML model was developed.
• The model was trained using in-silico generated rodent data (low-fidelity) and limited patient-specific human data (high-fidelity).
• Cardiac strain data was used as the sole input for predicting LGE ground truth.

Main Results:

• The multi-fidelity ML model demonstrated remarkable performance in predicting LGE ground truth.
• The study successfully identified infarct region location and extent non-invasively.
• The approach validates the augmentation of synthetic data with limited in vivo human data.

Conclusions:

• A novel, non-invasive ML-based methodology can accurately assess myocardial infarcts using only cardiac strain.
• This approach offers a safer alternative to gadolinium-based contrast agents.
• The findings present a new paradigm for developing prognostic tools in data-limited biomedical challenges.