Prediction-Conditioned RAG: A Data-Driven Digital Twin Approach for Diabetes Clinical Decision Support in Resource-Constrained Settings

Abstract

Diabetes management in Indonesia remains largely reactive, constrained by costly continuous glucose monitoring (CGM), uneven electronic health record (EHR) adoption, and reliance on manual self-monitoring logbooks. State-of-the-art digital twin frameworks for glycemic management achieve strong predictive accuracy but require infrastructure—Personal Health Knowledge Graphs, HL7 FHIR interoperability, and structured retrieval—that is infeasible in resource-constrained primary care. This paper proposes Prediction-Conditioned Retrieval-Augmented Generation (RAG), which conditions clinical document retrieval on a patient's predicted future glucose trajectory rather than the current state alone. The system integrates a simplified, data-driven patient digital twin built from logbook variables with a Random Forest glucose prediction module and a retrieval pipeline grounded in clinical guidelines. Consistent with scoping-review evidence that most claimed clinical digital twins lack full bidirectional synchronization, we position the system as a data-driven monitoring digital twin, with recommendations routed through doctor-mediated validation before reaching patients. On the OhioT1DM type 1 diabetes benchmark, evaluated at its native CGM resolution as a proof of concept, the Random Forest module attains an RMSE of 22.60 mg/dL and places 94.35% of predictions within Clarke Error Grid Zones A+B at a 30-minute horizon, comparable to a stacked LSTM baseline (22.04 mg/dL) at lower computational cost. In a small retrieval ablation over six clinically divergent scenarios (n=6), Prediction-Conditioned RAG targeted the correct anticipated condition in 100% of cases versus 0% for a current-state baseline, improving mean reciprocal rank from 0.417 to 0.666. We discuss surrogate-dataset limitations, the absence of clinical validation, and directions for larger-scale evaluation.