Human-in-the-Loop AI Digital Twin to Extend Virtual Precision Diabetes Care Between Visits: Results From a Randomized Clinical Trial

Background Type 2 diabetes (T2D) requires sustained lifestyle support, yet health systems face workforce and visit-capacity constraints that limit ongoing coaching between encounters. We evaluated a digitally enabled “human-in-the-loop” care support model that uses a predictive AI digital twin to generate daily, personalized short message service (SMS) feedback aimed at improving lifestyle adherence while maintaining glycemic stability. Methods The parent trial enrolled 40 adults (≥18 years) with T2D in a 6-month randomized trial. Participants completed 3 months of baseline observation followed by a 3-month intervention period (6,467 longitudinal data points; mean follow-up: 174 days). For the ancillary AI intervention, a subset of participants was randomized to either receiving AI-generated daily individualized feedback or no daily feedback. The online human-in-the-loop predictive control (OHLC) model used a transfer-learning artificial neural network digital twin trained on self-monitoring data (weight, food logs, physical activity, glucose). A particle swarm optimization controller identified actionable behavior changes aligned with glucose and weight goals; the digital twin was retrained weekly to incorporate newly accrued data. Results The OHLC model achieved ≥80% prediction accuracy across all diet-condition subgroups. During the 3-month AI intervention period, participants receiving AI-generated feedback exhibited a trend of improved daily steps and stable or improved total caloric and total carb intake aligned with dietary targets. Mixed ANOVA results showed that the AI intervention group exhibited significant weight loss compared with the control group (average loss:5.871 lbs.; p<.0001), while maintaining stable glucose levels throughout the study period (p= 0.661). In contrast, the non-AI group exhibited more modest weight reduction (average loss: 3.574 lbs.). Conclusions This study suggests that a human-in-the-loop AI digital twin model can support weight management without compromising glycemic stability, offering a potentially scalable approach for extending precision diabetes self-management support beyond clinic visits. Larger, multi-site studies should assess implementation outcomes, access, and system-level effects on utilization, cost, and quality metrics.