Researchers have developed an automated system for people with type 1 diabetes. The AI-based system provides advice to help these individuals avoid dangerously low blood glucose levels.

Man measuring glucose levelsShare on Pinterest
Could AI help people manage type 1 diabetes?

A preliminary study suggests that the system’s weekly recommendations on insulin doses and diet closely match those that diabetes specialists provide.

People with type 1 diabetes produce insufficient insulin, the hormone that the body uses to regulate blood glucose levels.

To maintain optimal glucose levels and avoid episodes of dangerously low or high blood sugar levels — known as hypoglycemia and hyperglycemia, respectively — people with this condition must take carefully controlled doses of insulin.

Many people with type 1 diabetes manage their condition successfully using a dosing regimen known as multiple daily injections, which involves injecting a long acting form of insulin once or twice a day, plus fast acting insulin at each mealtime.

In between mealtimes, they also have the option of injecting “correction doses” of fast acting insulin if their blood sugar levels rise too high.

However, repeated dosing errors over time increase a person’s risk of progressive damage to their eyesight, nervous system, and kidneys, and an acute episode of hypoglycemia can lead to coma or even death.

Diabetes specialists at Oregon Health & Science University (OHSU) in Portland say that several factors can lead to people giving themselves too much or too little insulin.

These factors include difficulty calculating doses, fears about overdosing, and changes in the body’s insulin sensitivity during exercise, illness, stress, and menstruation.

Endocrinologists (doctors who specialize in hormone disorders) can offer advice on any adjustments that a person needs to make to their dosing regimen and diet, but people may go for several months without an appointment.

To address this problem, researchers at OHSU used artificial intelligence (AI) to develop an algorithm that gives people weekly guidance based on data from a continuous glucose monitor, insulin pens for injecting insulin, and a wearable device that monitors physical activity.

In its final version, the algorithm issues its advice via a smartphone app called DailyDose.

The research appears in the journal Nature Metabolism.

To train their AI algorithm to issue advice, the researchers used virtual patients — mathematical representations of how a real person’s metabolism responds to food, insulin injections, and exercise.

To check that the resulting algorithm’s recommendations were accurate and safe, they fed it data from 25 real patients who were receiving treatment at OHSU. They then asked a panel of endocrinologists to review the same data and issue their advice.

The researchers report that recommendations from the algorithm tallied with those from the endocrinologists 67.9% of the time.

For comparison, they cite evidence suggesting that endocrinologists fully agree with each other about insulin advice to patients only about 41% of the time.

“Our system design is unique,” says lead author Nichole Tyler, a medical and doctoral student in the OHSU School of Medicine. “We designed the AI algorithm entirely using a mathematical simulator, and yet, when the algorithm was validated on real-world data from people with type 1 diabetes at OHSU, it generated recommendations that were highly similar to recommendations from endocrinologists.”

Based on almost 100 weeks of patient data, the endocrinologists also judged the algorithm’s advice to be safe more than 99% of the time.

Confident that their new algorithm was effective and safe, the researchers conducted a small clinical study involving 16 adults with type 1 diabetes.

Once a week for 4 weeks, a doctor reviewed the subjects’ glucose history together with recommendations from the algorithm, before issuing advice on insulin dosing and behavioral changes.

While the trial was too small to provide definitive results, the researchers report a reduction in the number of hypoglycemic episodes among the participants.

To develop the algorithm, the OHSU Harold Schnitzer Diabetes Health Center collaborated with the Artificial Intelligence for Medical Systems Lab that Peter Jacobs, Ph.D., associate professor of biomedical engineering in the OHSU School of Medicine, leads.

“There are other published algorithms on this, but not a lot of clinical studies. Very few have shown a statistically relevant outcome — and most do not compare algorithm recommendations with those of a physician. In addition to showing improvement in glucose control, our algorithm generated recommendations that had very high correlation with physician recommendations, with over 99% of the algorithm’s recommendations delivered across 100 weeks of patient testing considered safe by physicians.”

– Peter Jacobs, OHSU School of Medicine.

The clinicians and researchers at OHSU now plan to run several larger clinical trials of their DailyDose app spanning 8 and then 12 weeks.

The team will also compare the app with other insulin treatment strategies, including fully automated insulin delivery.