A novel deep learning system for STEMI prognostic prediction from multi-sequence cardiac magnetic resonance

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Imaging

Background:

• ST-elevation myocardial infarction (STEMI) poses significant cardiovascular risks.
• Current risk scores and imaging biomarkers have limited accuracy for predicting post-STEMI outcomes.
• Accurate early risk stratification is crucial for personalized treatment strategies.

Purpose of the Study:

• To develop and validate DeepSTEMI, an AI system for predicting 2-year major adverse cardiovascular events (MACE) after STEMI.
• To integrate multi-sequence cardiac magnetic resonance (CMR) images with clinical data for enhanced risk prediction.
• To compare DeepSTEMI's performance against existing clinical risk scores and manual imaging biomarkers.

Main Methods:

• Developed an end-to-end deep learning system (DeepSTEMI) with U-Net for segmentation and Transformer for prediction.
• Utilized a multicenter dataset (n=610) from the EARLY-MYO-CMR registry for development.
• Externally validated the system in 334 patients from three independent cardiac centers.

Main Results:

• DeepSTEMI demonstrated superior predictive performance in external validation (AUC 0.894, accuracy 94.3%).
• The model identified high-risk patients with a 20-fold increased MACE risk.
• SHAP analysis confirmed clinical-imaging synergy, and DeepSTEMI outperformed the Eitel score across subgroups.

Conclusions:

• DeepSTEMI offers an automated, scalable, and interpretable solution for post-STEMI risk stratification.
• The system advances cardiovascular risk prediction beyond current limitations.
• DeepSTEMI shows particular benefit in women and patients imaged 4-7 days post-STEMI.