The same types of machine-learning techniques used to pilot self-driving cars and beat top chess players could help keep blood sugar levels in safe limits in people with type 1 diabetes.
Scientists at the University of Bristol have found that reinforcement learning, a form of machine learning in which a computer program learns to make decisions by trying different actions, is significantly safer and more effective than commercially available blood glucose controllers. demonstrated to be excellent. By using offline reinforcement learning, where the algorithm learns from patient records, the researchers improved on previous research, showing that good glycemic control can be achieved by learning from patient decisions rather than trial and error. rice field.
Type 1 diabetes is one of the most prevalent autoimmune diseases in the UK and is characterized by a deficiency of insulin, the hormone responsible for blood sugar regulation.
Because many factors affect a person’s blood glucose levels, choosing the correct insulin dose for a given scenario can be a difficult and tedious task. Current artificial pancreas devices offer automated insulin administration, but are limited by their simple decision-making algorithms.
But today, a new study was published. Journal of Biomedical Informatics, We show that offline reinforcement learning could be an important milestone in caring for people living with this condition. The greatest improvement was in children, who had an additional 1.5 hours of increase within their target blood glucose range per day.
Children represent a particularly important group as they often cannot manage diabetes without assistance and improvements of this magnitude can significantly improve long-term health outcomes.
“My research explores whether reinforcement learning can be used to develop safer and more effective insulin administration strategies,” said lead author Harry Emerson, Department of Mathematics, Bristol School of Engineering.
“These machine learning-driven algorithms have demonstrated superhuman performance in a game of chess and piloting a self-driving car, thus performing highly personalized insulin administration from pre-collected blood glucose data. You may learn how.
“This particular work focuses specifically on offline reinforcement learning, where algorithms learn behaviors by observing good and bad examples of glycemic control.
“Traditional reinforcement learning methods in this area primarily rely on a trial-and-error process to identify appropriate behaviors, which can expose real-world patients to dangerous insulin dosing.”
Due to the high risk of misadministration of insulin, the experiment was performed using an FDA-approved UVA/Padova simulator. This simulator creates a series of virtual patients for testing type 1 diabetes control algorithms. A state-of-the-art offline reinforcement learning algorithm was evaluated against his one of the most widely used artificial pancreas control algorithms. This comparison was conducted on his 30 virtual patients (adults, adolescents, and children), 7,000 days of data were reviewed, and performance was assessed according to current clinical guidelines. The simulator was also extended to consider realistic implementation challenges such as measurement errors, inaccurate patient information, and limited amount of available data.
This study provides a basis for continued reinforcement learning research in glucose regulation. We demonstrate the potential of this approach to improve health outcomes in patients with type 1 diabetes, while highlighting the shortcomings of this technique and areas for future development.
The researchers’ ultimate goal is to introduce reinforcement learning into a real-world artificial pancreas system. Because these devices operate with limited patient monitoring, significant evidence of safety and efficacy is required for regulatory approval.
Dr. Harry added, “This study demonstrates the potential of machine learning to learn effective insulin dosing strategies from pre-collected type 1 diabetes data. The method investigated is the most widely used. It outperforms one of the artificial pancreas algorithms on the market today and demonstrates its ability to leverage human habits and habits.” Plan your schedule so that you can respond more quickly to dangerous events. “
