Machine Learning Method Holds Potential to Improve Lives of Type-1 Diabetes Patients

The same machine learning methods used for self-driving cars and chess-playing programs can be used to improve the lives of patients with type-1 diabetes. Researchers at the University of Bristol utilized reinforcement learning, a type of machine learning in which a computer program learns by trying different actions, to improve blood glucose control. The research shows that learning from patient records can help achieve better blood glucose control than trial and error. Currently, artificial pancreas devices use simplistic decision-making algorithms to provide automated insulin dosing, but are limited. Research shows that reinforcement learning, which has demonstrated superhuman performance in chess and self-driving cars, could learn from pre-collected blood glucose data to provide safe and effective insulin dosing for patients. This method could notably benefit children, who often rely on assistance in managing their diabetes, and experience a marked improvement in long-term health outcomes. The researchers’ ultimate goal is to use reinforcement learning in real-world artificial pancreas systems, but achieving regulatory approval will require significant evidence of safety and effectiveness.