Modeling Diabetes Risk and Progression With Public Health Data: Ontology-Guided, Simulation-Capable Digital Twin...

Area of Science:

• Digital Health
• Computational Biology
• Health Informatics

Background:

• Digital twins (DTs) offer potential for personalized healthcare but are limited by data scarcity and lack of standardization.
• Existing DT studies often use narrow or proprietary datasets, hindering generalizability.
• Public health datasets are rich but underutilized due to complexity and lack of semantic integration.

Purpose of the Study:

• Develop and evaluate an ontology-guided, agent-orchestrated framework for building simulation-capable DTs from public health data.
• Utilize large language models (LLMs) for semantic reasoning to support explainable feature structuring and risk prediction.
• Enable predictive "what-if" progression analysis for chronic diseases like diabetes.

Main Methods:

• Applied a 6-stage DT framework to the Midlife in the United States study data (waves 2 and 3).
• Used ontology- and LLM-assisted selection to identify 200 key predictors across biological, behavioral, psychosocial, and socioeconomic domains.
• Trained predictive models (random forest, XGBoost, logistic regression) and implemented a state-transition simulator for progression modeling and scenario analysis.

Main Results:

• Selected 200 relevant predictors from 9976 variables using ontology and LLM guidance.
• Achieved strong predictive performance for diabetes onset (e.g., Random Forest AUC=0.82).
• Simulator showed realistic risk state transitions (92.5% next-state prediction accuracy) and demonstrated potential impact of interventions like weight reduction.

Conclusions:

• Presents a foundational framework for creating progression-aware DTs from public datasets.
• Transforms static population data into interpretable, longitudinal health trajectory representations.
• Demonstrates the utility of public health data for robust, explainable DT models for risk analysis and hypothesis generation.