A methodological approach for creating virtual patient cohorts reflecting real-world diabetes treatment outcomes

Virtual patients (VPs) are widely used to evaluate the performance, scope, and robustness of glycemic-control strategies. However, existing virtual cohorts often include too few subjects or fail to reflect the diversity of the intended clinical population. Generating new VPs that reproduce real-patient (RP) characteristics is therefore crucial for advancing diabetes therapies. We present a data-driven method to construct VP cohorts matched to individual RPs using routine therapy and outcome data. Starting from published probability distributions of Hovorka model parameters, we generated candidate VPs by Monte Carlo sampling and retained only physiologically plausible parameter sets. For each RP, the common pool of physiologically plausible VPs was evaluated separately. VPs passing the RP-specific basal-insulin prefilter were then simulated under meal-and-exercise scenarios derived from that RP’s data. We then used a constraint satisfaction problem (CSP) to select, for each RP, the strictest similarity thresholds that preserved at least 20 matched VPs. Similarity was assessed using therapy parameters and CGM-derived outcomes. The method was tested using anonymized data from eight adults with type 1 diabetes (T1D) who participated in a clinical trial conducted at the Hospital Clínic of Barcelona. From 20,000 candidate VPs, 8,387 passed the physiological plausibility screening. The CSP-based filtering procedure retained 20–43 matched VPs per RP. The selected cohorts were characterized at the patient level using therapy and CGM-derived metrics, including basal insulin, TIR, hypoglycemia, severe hypoglycemia, hyperglycemia, and severe hyperglycemia. Across five protocol-matched in silico scenarios, 43 of 45 endpoint-by-scenario comparisons did not reach nominal significance in paired Wilcoxon signed-rank tests. Two comparisons in Scenario 2 reached nominal significance: severe hyperglycemia and glucose coefficient of variation. These findings support the feasibility of constructing VP cohorts matched to individual RP profiles, reproducing key therapy and CGM features under protocol-matched conditions. The approach may support preclinical controller tuning and robustness assessment, although further validation in larger and more heterogeneous datasets is required.