MACE Risk Prediction in ARVC Patients via CMR: A Three-Tier Spatiotemporal Transformer With Pericardial Adipose...

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Major adverse cardiac events (MACE) present a significant threat to patients with arrhythmogenic right ventricular cardiomyopathy (ARVC).
• Cardiac magnetic resonance (CMR) aids in MACE risk assessment, but challenges include small ARVC datasets and overlapping image features between MACE and non-MACE patients.
• Leveraging CMR's dynamic and spatial information is crucial for accurate MACE prediction in ARVC.

Purpose of the Study:

• To develop a deep learning model, the Three-Tier Spatiotemporal Transformer (TTST), for predicting MACE in ARVC patients.
• To address limitations of small datasets and overlapping image distributions in CMR-based MACE prediction.
• To enhance risk prediction by integrating pericardial adipose tissue (PAT) dynamics and positional information.

Main Methods:

• A novel Three-Tier Spatiotemporal Transformer (TTST) model was designed to extract and fuse features from 2D spatial, temporal, and inter-slice depth dimensions of CMR data.
• A pericardial adipose tissue (PAT) embedding unit was introduced to incorporate biomarker information, reducing reliance on large datasets.
• A patch voting unit was utilized to identify indicative cardiac regions, guided by PAT embedding.

Main Results:

• The TTST model demonstrated superior performance compared to existing methods in MACE prediction, achieving high AUC and accuracy on both internal and external datasets.
• TTST achieved a C-index of 0.744 independently and improved the existing 5-year risk score model's C-index from 0.686 to 0.777.
• Experimental results confirm TTST's effectiveness in leveraging limited CMR data for robust MACE risk prediction.

Conclusions:

• The TTST model offers a powerful deep learning approach for MACE risk prediction in ARVC patients using CMR.
• Incorporating PAT information and spatiotemporal features significantly enhances prediction accuracy, especially with limited datasets.
• TTST shows clinical utility, improving risk stratification when used alone or combined with existing models.