Multimodal graph learning for chagas disease classification

Area of Science:

• Medical Diagnostics
• Computational Biology
• Machine Learning

Background:

• Accurate Chagas disease staging is crucial for treatment but challenged by limited data and complex diagnostic sources.
• Trypanosoma cruzi infection presents multimodal diagnostic data (ECG, Echocardiogram, Doppler, ELISA) that are difficult to integrate effectively.

Purpose of the Study:

• To develop and validate a novel Graph Attention Network (GAT) architecture for enhanced classification of Chagas disease stages.
• To leverage multimodal data fusion and data augmentation for improved diagnostic accuracy and generalization in experimental models.

Main Methods:

• Implementation of a Graph Attention Network (GAT) architecture on fully connected graphs to integrate multimodal features.
• Utilizing a Variational Graph Autoencoder for data augmentation, reducing overfitting and enhancing generalization by creating synthetic subjects.
• Employing GNNExplainer for model interpretability to identify key clinical features influencing classification.

Main Results:

• The proposed GAT methodology achieved 100% Area Under the Receiver Operating Characteristic Curve (AUROC) in identifying Chagas disease infection stages.
• Significant performance optimization was observed in complex modalities like Doppler, demonstrating the model's effectiveness.
• The GAT model proved robust, with validated pathophysiological consistency and identification of important clinical features.

Conclusions:

• Graph Attention Networks offer a robust and effective method for classifying Trypanosoma cruzi infection stages by fusing multimodal data.
• The developed data augmentation strategy enhances model generalization and performance, addressing data scarcity challenges.
• The study validates the clinical relevance and interpretability of GATs in Chagas disease diagnostics.