Researchers have developed a data-driven method to create personalised virtual cohorts for type 1 diabetes, promising improved testing of glucose-control systems and personalised care strategies.
Researchers in Scientific Reports have described a data-driven way to build virtual patient cohorts that are closely matched to individual people with type 1 diabetes, a step they say could improve the testing of glucose-control systems before they are tried in the clinic. The study used anonymised data from eight adults treated at Hospital Clínic de Barcelona and aimed to address a long-running problem in diabetes simulation: many virtual cohorts are either too small or too narrow to reflect the range seen in real patients.
Virtual patients have been used for years to assess glucose-management strategies, including in hospital and intensive care settings. Earlier work published in Critical Care and in Artificial Intelligence in Medicine showed that model-based virtual trials and patient-derived simulations can help researchers examine tight glycaemic control protocols and decision-support tools without exposing patients to risk. More recently, studies have extended the idea to virtual care services and digital diabetes programmes, underlining the wider interest in virtual methods across diabetes management.
In the new Scientific Reports paper, the team started with published probability distributions for Hovorka model parameters, then used Monte Carlo sampling to generate 20,000 candidate virtual patients. After removing physiologically implausible profiles, 8,387 remained. Each real patient was then compared with the shared pool of virtual patients, with further filtering based on basal insulin and on meal-and-exercise scenarios drawn from that person’s own data. A constraint satisfaction approach was used to choose the tightest similarity thresholds that still left at least 20 matches per real patient.
The resulting matched cohorts contained between 20 and 43 virtual patients for each real patient. The researchers then compared therapy and continuous glucose monitoring measures such as basal insulin, time in range, hypoglycaemia, hyperglycaemia and glucose variability across five protocol-matched in silico scenarios. They reported that 43 of 45 endpoint-by-scenario comparisons were not significantly different in paired Wilcoxon signed-rank testing, although two measures in one scenario , severe hyperglycaemia and glucose coefficient of variation , did reach nominal significance.
The authors say the findings support the feasibility of building virtual cohorts tailored to individual real patients while preserving key treatment and glucose patterns under matched conditions. They argue that such cohorts could be useful for preclinical controller tuning and for testing how robust a glucose-control strategy may be across different patient profiles. Even so, they caution that the approach still needs validation in larger and more diverse datasets before it can be considered broadly representative.
Disclaimer: This content is for informational purposes only and is not intended to be a substitute for professional medical judgment, advice, diagnosis, or treatment.
