AEM: An interpretable multi-task multi-modal framework for cardiac disease prediction.

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Cardiovascular disease (CVD) is a leading global cause of mortality.
• Early heart failure (HF) prediction is challenging due to symptom heterogeneity.
• A multidisciplinary approach is needed for comprehensive cardiac evaluation.

Purpose of the Study:

• To develop a novel pre-training framework, the Anatomy-Electrocardiogram Model (AEM), to analyze interactions between 3D cardiac anatomy and electrocardiogram (ECG).
• To improve the early prediction of heart failure and survival analysis using multi-modal data.

Main Methods:

• AEM utilizes a multi-task self-supervised learning scheme with masked reconstruction and cardiac measurement regression.
• The model integrates background-free 3D cardiac anatomy (point clouds) with 12-lead ECG data.
• Experiments were conducted on multi-modal datasets from the UK Biobank.

Main Results:

• AEM achieved an Area Under the Receiver Operating Characteristic Curve of 0.8192 for incident HF prediction.
• The model obtained a concordance index of 0.6976 for survival prediction.
• AEM outperformed state-of-the-art multi-modal methods and demonstrated interpretability by recognizing clinically plausible patterns.

Conclusions:

• The AEM framework effectively integrates 3D cardiac anatomy and ECG data for enhanced cardiac state evaluation.
• The model shows significant potential for improving early heart failure prediction and survival analysis.
• AEM's interpretability suggests strong clinical relevance and association with known disease features.