AI enhances computational performance of artificial pancreas system

A research team from the University of Virginia (UVA) Center for Diabetes Technology has demonstrated that incorporating AI into an artificial pancreas system for regulating type 1 diabetes improves the system’s efficiency without sacrificing performance.

The researchers emphasized that automated insulin delivery (AID) systems play a key role in the management of type 1 diabetes for many patients, as they automatically monitor and regulate blood sugar.

However, the computational demands of these systems are higher than those of traditional diabetes management tools, like insulin pumps, making them more challenging to utilize in real-world settings.

To tackle this challenge, the research team set out to determine whether the incorporation of AI could improve efficiency.

The researchers tested an AI-based artificial pancreas, referred to as a “Neural Net Artificial Pancreas” (NAP), against a traditional artificial pancreas in a cohort of 15 patients with type 1 diabetes. Each participant was randomly assigned either the NAP or the traditional system and tasked with following their regular routines as closely as possible for 20 hours while using the device.

The NAP system successfully kept patients’ blood sugar in the desired range 86 percent of the time compared to the traditional system’s 87 percent, but the AI assistance enabled significantly improved efficiency, reducing computational demands six-fold over its counterpart.

These findings led the researchers to conclude that NAP could allow developers to incorporate the system into devices with low computational power – such as insulin pumps or pods – to increase efficiency.

“Neural-net implementation allows the algorithm to learn from the data of the person wearing the system,” explained Boris Kovatchev, PhD, director of the UVA Center for Diabetes Technology, in the news release. “This opens the door to real-time, AI-driven personalized insulin delivery.”

However, the small participant cohort warrants further investigation into the tool, but the research team underscored that an advanced version of NAP could potentially adapt and improve its own performance over time using data from thousands of users, rather than information from just one individual.

The research represents a first in the realm of AI-assisted AID devices.

“So far, this is the first clinical trial of a data-driven artificial pancreas system, which used [an] extensively trained neural network to deliver insulin automatically," said Kovatchev.

This research is the latest to assess how AI could transform diabetes management.

In August, a research team from Emory University detailed the validation of a deep learning model capable of flagging early signs of diabetes by analyzing EHR data and routinely collected chest radiographs.

The researchers emphasized that current screening guidelines for diabetes may miss a large number of at-risk patients, leading the team to investigate the potential of AI-based risk stratification.

The analysis revealed that the location of fatty tissue visible in chest X-rays is key for predicting diabetes risk, allowing the model to successfully identify high-risk patients up to three years before an official diagnosis.