Retrieval-Augmented Simulation for Interpreting Probabilistic Digital Twins in Type 1 Diabetes

Personalized digital twins for type 1 diabetes (T1D) can simulate future glucose responses under different therapeutic conditions, but their probabilistic outputs require structured interpretation before supporting clinician-facing reasoning. We introduce retrieval-augmented simulation (RAS), a modular framework connecting probabilistic digital-twin simulation with guideline-grounded report generation.

Retrieval-augmented simulation summarizes simulated glucose trajectories, converts them into structured clinical descriptors, retrieves relevant guideline evidence, and generates uncertainty-aware reports for clinician review. We evaluated RAS using data from 35 individuals with T1D in a controlled ablation study involving 2 locally hosted large language models (LLMs) (LLaMA 3.1:8b and Qwen 3:8b). Two independent LLM-based evaluators, GPT-5.5 Thinking and Kimi 2.6 Thinking, assessed report quality using a standardized rubric.

Structured clinical descriptors were essential for guideline selection: their removal substantially altered the retrieved evidence, demonstrating that qualitative clinical framing influenced evidence selection more strongly than numerical values alone. For LLaMA, the full RAS pipeline achieved the strongest overall performance, maintaining high numerical faithfulness (NumCov: $0.98\pm 0.03$) and consistent guideline coupling. In contrast, Qwen exhibited substantial model-specific safety risks and frequently distorted simulation-derived numerical values (NumCov: [Formula: see text]). Directional bias analysis showed that Qwen systematically under-reported time in range by a mean of $-10.6$ percentage points and over-reported time below range, thereby inflating hypoglycemia risk. These safety findings were consistent across both independent evaluators, with a mean score difference of $>2.1$ points.

Retrieval-augmented simulation provides a reproducible framework for the expert-supervised interpretation of digital-twin simulations. However, aggregate framework scores alone are insufficient for clinical validation. Downstream language models must be rigorously evaluated for numerical faithfulness and directional bias to prevent unsafe, hallucinated clinical narratives.