Auxiliary-conditioned cross-attention with physiologically interpretable features for chagas disease detection from...

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Chagas disease poses a significant public health risk in endemic areas.
• Chronic Chagas disease frequently leads to cardiac conduction abnormalities detected via electrocardiograms (ECGs).
• Automated ECG screening for Chagas disease is hindered by diverse datasets and uncertain labels.

Purpose of the Study:

• To develop a robust automated screening framework for Chagas disease cardiac involvement.
• To address challenges posed by dataset heterogeneity and label uncertainty in ECG analysis.
• To integrate diverse data sources, including physiological features and demographics, for improved screening.

Main Methods:

• A hybrid deep learning architecture combining a 1D ResNet encoder and a bidirectional GRU was proposed.
• An auxiliary-conditioned cross-attention module integrated handcrafted physiological features and demographics.
• A source-aware weighted binary cross-entropy objective was employed to manage data source reliability.

Main Results:

• The proposed framework achieved a score of 0.347 on the REDS-II leaderboard-validation set and 0.218 on the hidden test set in the PhysioNet/Computing in Cardiology Challenge 2025.
• The team ranked 17th out of 41 eligible teams, demonstrating competitive performance.
• The model's performance highlights its potential in a challenging competition setting.

Conclusions:

• Conditioning cardiac abnormality detection on interpretable QRS and conduction descriptors enhances screening transparency and physiological grounding.
• The study underscores the inherent difficulties in achieving generalization across heterogeneous patient cohorts.
• The developed framework offers a physiologically informed approach to Chagas disease screening.