JMIR Publications

ABSTRACT

Background:

Cardiovascular diseases (CVDs) are the leading global cause of death. Electrocardiograms (ECGs) are essential for cardiac screening but face limitations in traditional interpretation methods, particularly in detecting complex or comorbid conditions.

Objective:

This study aimed to develop and evaluate CaMPNet, a transformer-based multimodal deep learning model for automated detection of multiple cardiovascular comorbidities using raw ECG waveforms, structured ECG features, and demographic data.

Methods:

CaMPNet integrates 12-lead raw ECG signals, structured ECG metrics (e.g., PR/QRS/QT intervals), and patient demographics via cross-attention fusion. The model was trained on 384,877 ECG records from the MIMIC-IV-ECG dataset and evaluated across 12 cardiovascular conditions. Subgroup analyses were conducted by age and sex. Model performance was benchmarked against ResNet-based and single-modality baselines.

Results:

CaMPNet achieved a mean AUC of 0.865 and AUPRC of 0.475, outperforming baseline models. Subgroup analysis demonstrated consistent performance across demographics. Attention maps provided clinically interpretable insights (e.g., ST-segment elevation in STEMI). Ablation studies confirmed robustness to missing modalities.

Conclusions:

CaMPNet demonstrates strong multi-label classification performance and interpretability across diverse cardiovascular conditions. However, clinical deployment requires further validation via multi-center, prospective studies, and enhancements in rare-class detection and explainability. Future work will focus on improving generalizability, label quality, and real-world applicability. Clinical Trial: Not applicable.