Physiology-guided artificial intelligence digital twins for simulating glucose dynamics in type 2 diabetes

Abstract

Background: Metabolic simulators help accelerate the development of diabetes technologies, but existing mechanistic models for type 2 diabetes (T2D) are difficult to individualize and provide limited explanation of the complex interplay between physiology and available treatments. Methods: First, we trained NeuralOGTT, a neural network state-space model that replicates glucose-insulin dynamics during the oral glucose tolerance test (OGTT) in T2D. Next, we determined the minimal model architecture using a variance-based sensitivity analysis. Finally, we created digital twins (DT-NeuralOGTT) by adding auxiliary sub-networks to the NeuralOGTT trained using individuals data from a free-living observational study (N=38, age: 50 ± 10 years; 27 females; 1 high risk of diabetes, 9 pre-T2D, and 28 T2D). We simulated study scenarios and assessed the digital twins ability to replicate real-world glucose outcomes: time in range (TIR,$70-180$ mg/dL), time in tight range (TITR,$70-140$ mg/dL), time below range (TBR, < 70 mg/dL), and time above range (TAR, > 180 mg/dL). Results: NeuralOGTT replicated glucose dynamics during OGTTs. Personalized DT-NeuralOGTT enabled generalization under free-living conditions across 12-h simulations, achieving simulated glucose outcomes that are statistically identical to observed outcomes: TIR: 91.8 (95% CI:82.2–96.0)% vs. 92.1 (95% CI:83.5–95.8)%, TITR: 56.0 (95% CI:36.6–77.4)% vs. 59.3 (95% CI:48.7–76.4)%, TAR: 6.3 (95% CI:1.6–15.8)% vs. 7.4 (95% CI:3.8–15.7)%, and TBR: 0.5 (95% CI:0.0–1.0)% vs. 0.1 (95% CI:0.1–0.2)%. Conclusion: NeuralOGTT models glucose–insulin dynamics using a structure that preserves explainability while requiring small amounts of training data. DT-NeuralOGTT incorporates individual variability to better capture dynamics in pre-T2D and T2D. Code available: https://github.com/mosqueralopez/T2DSim_AI