Understanding Artificial Pancreas Systems: Results from a 24-Month Trial

Related and Discussions

Subscribe now for the latest breakthroughs, expert insights, and cutting-edge updates in diabetes care—delivered straight to your inbox.

Stay informed. Stay ahead.

The virtual trial showed GLP-1 reduced HbA1c more than insulin (1.21 mmol/mol difference) and placebo (2.58 mmol/mol difference), matching the real trial's finding that GLP-1 works best for blood sugar control. The AI correctly identified treatment rankings, though exact numbers differed because it tested drug classes rather than specific medications.

GLP-1 reduced BMI more than insulin (0.79 kg/m² difference) and placebo (0.61 kg/m² difference). This matches the real trial's weight loss findings, though the virtual trial measured BMI instead of actual weight. The results show GLP-1 helps with weight management beyond blood sugar control.

Virtual trials showed GLP-1 lowered blood pressure more than insulin (3 mmHg difference) and placebo (2 mmHg difference), matching the real trial's conclusions. The model captured GLP-1's heart health benefits beyond blood sugar control.

Insulin helped lower HbA1c but was less effective than GLP-1, matching the real trial. When tested in broader groups, some people responded better to insulin than GLP-1, showing treatment benefits can vary by individual.

Insulin was linked to BMI increase compared to GLP-1 (0.79 kg/m² difference), matching known effects that insulin can cause weight gain. This shows a common challenge with insulin treatment.

Insulin showed small blood pressure reduction but was less effective than GLP-1 (3 mmHg difference), supporting the heart health advantages of GLP-1 drugs over traditional insulin.

The virtual trial results align with findings from the original LEAD-5 study and other GLP-1 research. The LEAD-5 trial found liraglutide worked better than insulin glargine for blood sugar and weight. The LEADER trial showed liraglutide also benefits heart health, consistent with this study's blood pressure findings.

This study's main contribution is demonstrating a new AI method for testing treatments, not discovering new clinical effects. The finding that treatment benefits vary by individual aligns with growing interest in personalized medicine approaches. As regulators consider AI-generated evidence, this framework may help fill gaps when traditional trials aren't practical.

This study gets a Moderate rating because the AI model successfully matched the results from a real trial. The model used data from over 5,000 people and correctly identified which treatment worked best. The findings were consistent across blood sugar, weight, and blood pressure.

However, this is a proof-of-concept study testing a new AI method. The model tested drug classes rather than specific medications, which may miss important differences. Removing incomplete data may have introduced bias. More validation is needed before this method can be used widely.

AI models can recreate results from real diabetes trials, correctly showing GLP-1 drugs work better than insulin for blood sugar, weight, and blood pressure.

Treatment benefits may vary by individual, with GLP-1 not always the best choice for everyone.

This AI method needs more testing before it can be used widely for treatment decisions.

This study shows AI models can match real trial results, potentially helping test treatments in groups typically excluded from trials.

The model correctly identified that GLP-1 works better than insulin across multiple outcomes.

Results are less reliable for predicting exact effect sizes because the model used drug classes rather than specific medications.

The finding that treatment benefits vary by person suggests one-size-fits-all approaches may miss important differences.

This is early-stage research that needs validation across more trials before clinical use.

The study tested three treatment groups to match the original LEAD-5 trial. The first group received GLP-1 drugs, which help the body release insulin, reduce appetite, and slow digestion. The second group received basal insulin, a long-acting insulin that controls blood sugar throughout the day. The third group received placebo medications that don't directly affect blood sugar.

All three groups continued taking metformin and sulfonylurea as background treatment. The AI model tested drug classes rather than specific medications to account for real-world prescribing patterns. In virtual trials, people were randomly assigned to treatments to eliminate bias, just like in real clinical trials.

GLP-1 receptor agonists (liraglutide), basal insulin analogs (insulin glargine), metformin, sulfonylureas, proton pump inhibitors (placebo proxy), statins (placebo proxy)

MacLellan CR, Petkov H, McKeag C, et al. Emulating real-world GLP-1 efficacy in type 2 diabetes through causal learning and virtual patients. PLOS Digit Health. 2025;4(7):e0000927. doi:10.1371/journal.pdig.0000927

The AI model successfully matched the original trial's results. GLP-1 drugs worked better than both basal insulin and placebo for all three outcomes. For HbA1c, GLP-1 reduced levels 1.21 mmol/mol more than insulin and 2.58 mmol/mol more than placebo. For BMI, GLP-1 reduced it 0.79 kg/m² more than insulin and 0.61 kg/m² more than placebo. For blood pressure, GLP-1 lowered it 3 mmHg more than insulin and 2 mmHg more than placebo.

When researchers tested the AI on broader groups beyond the original trial criteria, they found GLP-1 wasn't always the best choice. Some people responded better to insulin or even placebo. This suggests treatment effectiveness may depend on individual characteristics.

Tested drug classes, not specific medications

Virtual patients showed less variation than real people

AI models can recreate trial results, but need more validation before use in treatment decisions.

The study measured body mass index instead of weight, which may respond differently to treatment. The model tested drug classes rather than specific medications like liraglutide and insulin glargine, which could mask important differences. The AI assumed variables were independent when they may actually be related in real life. Using yearly averages may have missed short-term treatment effects. The study excluded all records with missing data, which could introduce bias. The virtual patients showed less variation than real people for some measures, suggesting the model may not fully capture real-world diversity.

HbA1c → ↓ (GLP-1 better than insulin)

Weight (BMI) → ↓ (GLP-1 better than insulin)

Blood pressure → ↓ (GLP-1 better than insulin)

Liraglutide vs insulin glargine and placebo in combination with metformin and sulfonylurea therapy in type 2 diabetes mellitus (LEAD-5 met+SU)

Liraglutide and Cardiovascular Outcomes in Type 2 Diabetes

Type 2 Diabetes - Diagnosis and Treatment

Researchers used AI to create virtual patients based on data from 5,476 people with type 2 diabetes in Scotland. The AI learned patterns of how different treatments affect blood sugar, weight, and blood pressure. It then created virtual versions of the LEAD-5 trial, which originally compared GLP-1 drugs to basal insulin and placebo.

Each virtual trial included 232 people per treatment group, matching the original trial's patient characteristics. Participants were middle-aged adults with an average age of 58, BMI of 30, and HbA1c around 67 mmol/mol. All were taking metformin and sulfonylurea. The model tracked outcomes over 12 months. Researchers used 60 different random starting points to ensure results were consistent.

AI virtual trials match real trial results for GLP-1 vs insulin in type 2 diabetes

Researchers used AI to create virtual patients and test whether they could recreate results from a real diabetes trial. They trained the AI using data from over 5,000 people with type 2 diabetes. The AI successfully matched the original trial's findings that GLP-1 drugs work better than basal insulin for blood sugar control, weight loss, and blood pressure.

The AI model correctly showed GLP-1 was more effective across all three outcomes tested. When researchers used the AI to test treatments in broader groups of people beyond the original trial, they found GLP-1 wasn't always the best choice for everyone. This suggests treatment benefits may vary depending on individual characteristics.

This tests an AI method, not actual treatments, so it doesn't provide new clinical evidence.

Virtual patients showed less variation than real people, which may miss how some individuals respond.

The model tested drug classes, not specific medications, so exact effects may differ.

AI model using data from 5,476 people to recreate a diabetes drug trial.

People with type 2 diabetes on metformin and sulfonylurea considering GLP-1 or insulin treatment; healthcare providers choosing between treatment options; researchers testing new AI methods for drug evaluation.

People with type 2 diabetes considering GLP-1 receptor agonist therapy, basal insulin therapy, or combination treatments; healthcare providers making treatment decisions for diverse patient populations; researchers and policymakers evaluating treatment effectiveness beyond traditional trial populations.

ADA Standards of Care in Diabetes - 2025

Researchers used AI to create virtual patients and test whether they could recreate results from a real diabetes trial. They trained the AI using data from over 5,000 people with type 2 diabetes. The AI successfully matched the original trial's findings that GLP-1 drugs work better than basal insulin for blood sugar control, weight loss, and blood pressure.
The AI model correctly showed GLP-1 was more effective across all three outcomes tested. When researchers used the AI to test treatments in broader groups of people beyond the original trial, they found GLP-1 wasn't always the best choice for everyone. This suggests treatment benefits may vary depending on individual characteristics.

AI virtual trials match real trial results for GLP-1 vs insulin in type 2 diabetes

Study at a glance

Main effects

Evidence Suggest

Who this applies to

Keep in Mind

Between the Lines

Unlock Full Analysis

Journal Reference

No ads. No tracking.

Secure & private

Always up to date