Explainable artificial intelligence and domain adaptation for predicting HIV infection with graph neural networks

Area of Science:

• Computational epidemiology
• Network science
• Machine learning

Background:

• HIV remains a significant public health concern, particularly among younger sexual minority men.
• Social network analysis offers insights into HIV transmission dynamics.
• Predictive modeling using complex network data presents challenges due to sparsity and heterogeneity.

Purpose of the Study:

• To investigate explainable deep learning methods, specifically graph neural networks (GNNs), for predicting HIV infections.
• To evaluate the transferability of these models across different datasets using domain adaptation.
• To identify key network features associated with HIV risk.

Main Methods:

• Utilized network data from two cohorts of younger sexual minority men (SMM) in Chicago and Houston.
• Employed graph attention network (GAT) models for prediction and GNNExplainer for feature importance.
• Implemented domain adaptation to assess model performance when transferring between city datasets.

Main Results:

• Domain adaptation improved GAT model prediction accuracy compared to single-city training.
• Feature importance analysis revealed consistent risk indicators across different cities.
• GAT models demonstrated potential for predicting HIV infections by leveraging social network information.

Conclusions:

• Graph attention network models can effectively address data sparsity in HIV research.
• These models offer a powerful approach for predicting individual HIV risk.
• Domain adaptation enhances the generalizability of GNNs for understanding HIV transmission patterns.