New AI Model from UTHealth Houston Researchers Predicts Individual Infection Risk Using Social Network Data

Researchers at UTHealth Houston have developed an artificial intelligence framework that predicts COVID-19 infection risk using complex social network data spanning personal, household, and community layers. The Multilayer Modular Fusion Graph Attention Network (MMF-GAT) was evaluated on COVID-19 surveillance data from the Houston Health Department.

The team introduced MMF-GAT, built on a graph attention network with fusion techniques and explainable artificial intelligence (XAI), in PLOS Complex Systems. The framework is a graph-based deep learning prediction tool intended to support targeted planning and resource allocation in public health and beyond.

Key Highlights of MMF-GAT:

• High Predictive Performance: Exceeded five standard models with Area Under the Curve (AUC) of 0.90, accuracy of 0.78, and F1 score of 0.72 on a real-world dataset of 2,264 individuals. The framework is built for uneven, real-world network patterns, where cases cluster in local hotspots within overlapping groups and spread via highly connected hubs linking otherwise separate communities.
• Explainable AI (XAI): In addition to generating predictions, the system highlights the contributing factors. For the evaluated dataset, important features included household size, number of personal contacts, and affiliations with education centers that may function as high-contact hubs.

• Advanced Architecture:A multilayer, modular, and interpretable design that preserves each layer’s unique structure and then fuses them for prediction, making it suitable for epidemic forecasting and adaptable to other multilayer network applications.

“Our new artificial intelligence architecture uses graph-based neural networks to bring together sociology, network science, and epidemiology. It captures patterns across overlapping social groups, positions within personal networks, and connections spanning different contexts. This interdisciplinary design provides a more comprehensive understanding of complex human connections,” said Kayo Fujimoto, PhD, Distinguished Professor at UTHealth School of Public Health and principal investigator on the project.

“Beyond public health, this framework has the potential to generalize to other complex networked systems, supporting node-level prediction across domains such as molecular biology, systems medicine, organizational behavior in healthcare, and consumer networks,” said Fujimoto.

The full manuscript is available online now. For more information on the MMF-GAT Framework, please contact Fujimoto at [email protected].