Researchers have developed a novel approach to construct virtual patient cohorts that better mirror real individuals with type 1 diabetes, potentially enhancing computer-based testing of glucose control systems and insulin management strategies.
Researchers at Scientific Reports, part of the Nature portfolio, have described a data-driven way to build virtual patient cohorts that more closely resemble real people with type 1 diabetes, a step that could make computer-based testing of glucose-control systems more reliable. The approach uses routine therapy and glucose data to match synthetic patients to individual real patients, rather than relying on small or overly uniform cohorts that may miss important clinical variation.
The study starts with published probability distributions from the Hovorka diabetes model, a widely used physiological framework for simulating glucose and insulin dynamics. The team generated 20,000 candidate virtual patients through Monte Carlo sampling, then removed parameter sets that were not physiologically plausible. From there, each real patient was compared with a shared pool of feasible virtual patients, with an initial screen based on basal insulin requirements before the candidates were tested under meal and exercise conditions drawn from that patient’s own record.
To choose the closest matches, the researchers used a constraint satisfaction method to set the strictest similarity thresholds that still preserved at least 20 matched virtual patients per real patient. Similarity was judged using treatment variables and continuous glucose monitoring measures, including time in range, hypoglycaemia, hyperglycaemia and glucose variability. In the anonymised test set of eight adults who had taken part in a trial at Hospital Clínic de Barcelona, 8,387 candidates passed the physiological screen, and the final matched cohorts contained between 20 and 43 virtual patients for each real patient.
Across five protocol-matched in silico scenarios, 43 of 45 endpoint comparisons were not statistically significant in paired Wilcoxon signed-rank testing, suggesting that the virtual cohorts reproduced the real patients’ key features closely under the tested conditions. Two comparisons in one scenario did show nominal significance: severe hyperglycaemia and glucose coefficient of variation. The authors say the method could help with preclinical tuning of automated insulin controllers and with testing how robust they are, but they also stress that larger and more diverse datasets will be needed before the approach can be considered broadly validated.
Disclaimer: This content is for informational purposes only and is not intended to be a substitute for professional medical judgment, advice, diagnosis, or treatment.
