Abstract Body: Introduction:
Obesity and cardiometabolic diseases are leading contributors to global morbidity and mortality. Their interdependent progression is not well captured by existing models, which often rely on oversimplified assumptions and lack the flexibility to evaluate diverse populations.
Hypothesis:
We assessed the hypothesis that a generative AI digital twin model could accurately simulate long-term, multivariate trajectories of obesity and cardiometabolic comorbidities, enabling evaluation of intervention effects across subgroups and clinical benefits associated with sustained weight loss.
Methods:
The Dynamic Evaluation of Cardiometabolic and Obesity DiseasE (DECODE) model was developed using a Conditional Restricted Boltzmann Machine (CRBM) architecture. Training data included adult patients (≥18 years) from a U.S. electronic health record database (2007–2024, Dandelion Health) and a U.S. commercial claims database (2016–2023). Inputs comprised >100 static and longitudinal variables (e.g., demographics, body mass index [BMI], comorbidities, medications). Validation compared observed versus synthetic distributions, variances, and correlation structures in a held-out cohort (≥10,000 patients) using Pearson correlations and Intersection over Union (IoU). Subgroup validation examined performance by baseline type 2 diabetes, age ≥65 years, BMI ≥40, sex, race, and GLP-1RA prescription. Simulations estimated the 5- and 10-year effects of sustained 10% weight loss on cardiometabolic outcomes (including heart failure and atrial fibrillation) and bariatric surgery.
Results:
The trained DECODE model achieved high validity and concordance with observed data distributions, variances, and correlation structures (ρ≥0.94; overall and subgroup-specific IoUs ranged 0.88-0.99 and 0.78-0.99, respectively). Sustained 10% weight loss was associated with notable reductions in the 5- and 10- year cumulative incidence (5-year risk ratio [RR]; 10-year RR) of heart failure (0.83; 0.74) and bariatric surgery (0.83; 0.88). The association with atrial fibrillation was minimal (0.98; 0.96).
Conclusions:
DECODE provides a dynamic, high-resolution framework for simulating obesity and cardiometabolic disease progression across diverse populations. Generative AI digital twins enable rigorous evaluation of long-term interventions, with simulations demonstrating significant reductions in cardiometabolic risk at both 5 and 10 years following sustained weight loss.