JMIR Publications

ABSTRACT

Background:

Digital twins (DTs) offer a transformative paradigm for healthcare by creating dynamic, individualized models that simulate disease trajectories and support personalized interventions. However, DT development remains limited by the scarcity of standardized, temporally structured, and multidomain data suitable for modeling chronic disease progression. Most existing DT studies rely on narrowly scoped or proprietary datasets, restricting generalizability. Public health datasets such as the Midlife in the United States (MIDUS) study provide rich biopsychosocial information but are underused due to their structural complexity and lack of semantic integration frameworks.

Objective:

This study aimed to design, implement, and evaluate a scalable, ontology-guided, and agentic-AI framework for constructing personalized, simulation-capable digital twins from large public health datasets. Using diabetes as a case study, the framework integrates multi-agent coordination, medical ontologies, and large language model (LLM) reasoning to enable explainable feature selection, risk prediction, and disease-progression simulation.

Methods:

A six-stage DT framework was developed and applied to MIDUS Wave 2 (baseline) and Wave 3 (follow-up) data. Ontology- and LLM-assisted feature selection identified predictors across biological, behavioral, psychosocial, and socioeconomic domains. Cleaned and harmonized data were used to train predictive models (Random Forest, XGBoost, Logistic Regression) to estimate diabetes onset at follow-up. A state-transition simulator was then implemented to model progression dynamics, quantify transitions across low-, medium-, and high-risk states, and evaluate counterfactual “what-if” interventions such as weight reduction and lifestyle improvement. Model performance was assessed using accuracy, F1-score, AUC, and calibration metrics.

Results:

From 9,976 candidate variables, ontology- and LLM-guided selection retained the top 200 most relevant predictors spanning biological, behavioral, psychosocial, and socioeconomic domains. Predictive modeling achieved strong discrimination, with Random Forest (AUC = 0.97, accuracy = 0.91) and XGBoost (AUC = 0.97, accuracy = 0.90) outperforming Logistic Regression (AUC = 0.94). The state-transition simulator reproduced realistic progression patterns: 33.9% of participants changed risk states between waves, and the high-risk group increased from 10.8% to 32.2%. Next-state prediction accuracy reached 92.5%, confirming the simulator’s ability to capture longitudinal dynamics. Counterfactual simulations demonstrated actionable outcomes: a uniform 10% weight reduction improved risk states for 6.7% of participants and reduced predicted diabetes incidence by 98 cases (576 → 478). A placebo test (0% weight change) produced < 0.3% difference in risk distribution, confirming model stability.

Conclusions:

This study introduces a generalizable, ontology-guided, and multi-agent framework for constructing personalized digital twins from public datasets. By combining semantic reasoning, multidomain predictors, and progression simulation, the framework transforms static population data into dynamic, interpretable representations of individual health trajectories. The proof-of-concept application to diabetes demonstrates that public health data can support robust, explainable, and intervention-aware digital twins for chronic disease prevention and management.