CARDIAC-FM: A Multimodal Foundation Model for Cardiovascular Risk Prediction Using ECG and Cardiac MRI.

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Imaging

Background:

• Atrial fibrillation and heart failure represent significant global health challenges.
• Current prediction models for these conditions often lack accuracy and broad applicability.
• There is a need for more robust and generalizable cardiovascular risk prediction tools.

Purpose of the Study:

• To develop CARDIAC-FM, a multimodal foundation model for enhanced cardiovascular risk prediction.
• To learn joint representations from electrocardiogram (ECG) and cardiac magnetic resonance imaging (MRI) data.
• To improve the accuracy and generalizability of predicting cardiovascular outcomes.

Main Methods:

• Developed CARDIAC-FM, a multimodal foundation model using contrastive learning.
• Trained the model on 57,609 paired ECG-cardiac MRI samples from the UK Biobank.
• Evaluated performance in two independent cohorts: the Cardiovascular Health Study (CHS) and the Multi-Ethnic Study of Atherosclerosis (MESA).

Main Results:

• CARDIAC-FM outperformed unimodal models in all evaluated cohorts.
• Combining ECG features with clinical risk scores provided additional predictive gains.
• The model demonstrated improved prediction for various cardiovascular outcomes with minimal fine-tuning.

Conclusions:

• Multimodal pre-training shows promise for generalizable cardiovascular risk prediction.
• CARDIAC-FM can predict cardiovascular risk using ECG alone or with risk scores, facilitating clinical use without MRI.
• The model's learned representations capture complementary cardiovascular risk dimensions.