Type 1 diabetes: an AI guides insulin settings as well as doctors

Why is it so hard to get insulin settings right?

In type 1 diabetes, the body no longer makes insulin, the hormone that lets sugar move from the blood into the cells. Many people use an insulin pump, which delivers it continuously, and a glucose sensor, which measures blood sugar all the time.

Despite these devices, most people with the condition do not reach their glucose goals. The authors point to a lack of expertise or a lack of time for doctors to analyse the complex data from the pump and the sensor.

Keeping glucose in the target range takes frequent adjustments, which depend on food, activity and, in young people, growth. That is exactly the kind of repetitive task where decision support could be useful.

How was the trial run?

The trial is called ADVICE4U. It lasted 6 months, across several centres in different countries. One hundred and eight young people aged 10 to 21, all using an insulin pump, were randomly split into two groups of 54.

In one group, insulin settings were adjusted remotely every three weeks on the advice of the AI system. In the other, they were adjusted with the help of doctors from academic centres specialised in diabetes. The aim was to check that the AI was no less effective and no less safe: this is called a non-inferiority trial.

What were the results?

The main outcome was time spent in the target glucose range, between 70 and 180 mg/dl (3.9 to 10.0 mmol/l). It was 50.2% in the AI group and 51.6% in the doctor group: the AI is therefore non-inferior.

For severe low blood sugar, the percentage of readings below 54 mg/dl was 1.3% with the AI and 1.0% with doctors, which remains non-inferior. Three severe diabetes-related events (two severe hypoglycaemias and one ketoacidosis, a serious complication caused by a lack of insulin) occurred in patients followed by doctors, and none in those followed by the AI.

In plain terms, the young people followed with the AI's help spent about as much time in range as those followed by doctors, with no extra safety signal in this trial.

What could this change for patients?

If the tool is confirmed, it could help care teams follow more patients, more often, with better-adapted settings, without everything depending on one doctor's time. That could matter for young people whose settings need regular review.

This is not an artificial pancreas acting alone: here, it is a system that suggests dose adjustments, remotely, every three weeks.

Doctors keep a central role: they set the frame of the treatment, and it is up to them to judge whether an AI suggestion suits the young person's situation.

What are the limits of this study?

The trial is modest in size: 108 young people. The AI was compared with doctors from specialist centres; we do not know what would happen against more ordinary follow-up. The outcome tested was non-inferiority: the AI does as well, not better.

The authors do not show an advantage over time beyond 6 months. Further studies, with more participants and other age groups, are needed before everyday use.

Finally, the study reports the average results of groups, so it cannot tell us how the AI would do for any one young person.