Type-1 diabetics may be able to maintain safe blood glucose levels with the same machine learning techniques that were used to pilot self-driving cars and defeat top chess players.
Researchers from the University of Bristol have demonstrated that commercial blood glucose controllers are significantly less safe and effective than reinforcement learning, a type of machine learning in which a computer program learns to make decisions by attempting various actions. The researchers show that good blood glucose control can be achieved by learning from the decisions made by the patient rather than by trial and error by using offline reinforcement learning, in which the algorithm learns from patient records. This improves on previous work and demonstrates this.
One of the most common auto-immune conditions in the UK is type 1 diabetes, which is caused by a lack of the hormone insulin, which controls blood glucose levels.
“My study looks into whether reinforcement learning can be used to develop safer and more effective insulin dosing strategies.”
Lead author Harry Emerson from Bristol’s Department of Engineering Mathematics,
The selection of the appropriate dose of insulin for a given situation can be difficult and burdensome due to the numerous factors that influence a person’s blood glucose levels. Automated insulin dosing is provided by current artificial pancreas devices, but their simplistic decision-making algorithms restrict their capabilities.
On the other hand, a brand-new study that was recently published in the Journal of Biomedical Informatics demonstrates that offline reinforcement learning might be a significant step in the care of people who have the condition. Children saw the greatest improvement, with an additional 1.5 hours in the desired glucose range per day.
Due to the fact that children are frequently unable to manage their diabetes on their own, they are a particularly significant demographic, and an increase of this size would significantly improve long-term health outcomes.
“My research explores whether reinforcement learning could be used to develop safer and more effective insulin dosing strategies,” stated lead author Harry Emerson of Bristol’s Department of Engineering and Mathematics.
“These machine learning-driven algorithms have demonstrated superhuman performance in chess and driving self-driving cars, so they might be able to learn to do highly personalized insulin dosing from blood glucose data that has already been collected.”
“Offline reinforcement learning, in which the algorithm learns to act by observing examples of good and bad blood glucose control, is the primary focus of this work. In this area, prior reinforcement learning methods mostly use trial and error to find good actions, which could put a real-world patient at risk of taking unsafe insulin doses.”
The FDA-approved UVA/Padova simulator, which creates a suite of virtual patients to test type 1 diabetes control algorithms, was used in the experiments because of the high risk of administering the wrong amount of insulin. One of the most widely used artificial pancreas control algorithms was compared to cutting-edge offline reinforcement learning algorithms. This comparison took into account 7,000 days of data from 30 virtual patients (adults, adolescents, and children), and performance was evaluated in accordance with current clinical guidelines. Additionally, the simulator was expanded to take into account realistic obstacles to implementation, such as measurement errors, incorrect patient information, and insufficient data.
The foundation for future reinforcement learning studies on glucose control is provided by this work, which highlights the method’s shortcomings and areas for necessary future development while simultaneously demonstrating the approach’s potential to improve the health outcomes of type 1 diabetes patients.
Implementing reinforcement learning in real-world artificial pancreas systems is the ultimate objective of the researchers. To receive regulatory approval, these devices will need substantial evidence of safety and effectiveness because they operate without patient oversight.
“This research demonstrates the potential of machine learning to learn effective insulin dosing strategies from pre-collected type 1 diabetes data,” Emerson added. One of the most widely used commercial artificial pancreas algorithms is outperformed by the investigated method, demonstrating that it can take advantage of a person’s routine and schedule to respond more quickly to dangerous situations.
More information: Harry Emerson et al, Offline reinforcement learning for safer blood glucose control in people with type 1 diabetes, Journal of Biomedical Informatics (2023). DOI: 10.1016/j.jbi.2023.104376
