Multimodal Transformer-Based Electrocardiogram Analysis for Cardiovascular Comorbidity Detection: Model Development...

Area of Science:

• Artificial Intelligence in Medicine
• Cardiology
• Machine Learning for Healthcare

Background:

• Cardiovascular diseases are a leading cause of mortality globally.
• Traditional electrocardiogram (ECG) interpretation faces challenges with subjective variability and limited sensitivity.
• Complex cardiovascular pathologies require more advanced diagnostic tools.

Purpose of the Study:

• To develop an advanced AI model for cardiovascular disease diagnosis using multimodal data.
• To introduce the Cardiovascular Multimodal Prediction Network (CaMPNet), a novel transformer-based architecture.
• To integrate raw ECG waveforms, structured ECG features, and demographic data for enhanced prediction.

Main Methods:

• CaMPNet utilizes a transformer-based multimodal architecture with cross-attention fusion.
• The model was trained on a large dataset (384,877 records) from the MIMIC-IV-eICU database.
• Evaluated across 12 cardiovascular disease labels with internal and temporal external validation.

Main Results:

• CaMPNet achieved a mean AUC of 0.845, outperforming baseline models and single-modality approaches.
• Consistent performance was observed across demographic subgroups.
• Temporal external validation showed moderate discriminative ability (AUC=0.715), with key diseases maintaining high AUCs.
• Attention visualization revealed clinically interpretable patterns, and ablation studies confirmed input tolerance.

Conclusions:

• CaMPNet provides a robust and interpretable AI framework for ECG-based cardiovascular diagnosis.
• The model demonstrates scalability for comorbidity screening and continual learning.
• CaMPNet addresses real-world temporal dynamics in healthcare data.