The number of people with diabetes continues to increase worldwide, especially in low-and middle-income countries. In 2012, diabetes accounted for the death of 1.5 million people. This disease can be categorized into type 1, 2, and gestational diabetes.
Type 1 diabetes is a chronic condition where a person's body produces a high blood glucose or sugar level. Also known as juvenile diabetes or insulin-dependent diabetes, this lifelong autoimmune disease prevents a person's pancreas from making insulin.
Challenges in Managing Type 1 Diabetes
People with type 1 diabetes produce very little or no insulin on their own, so they need to track the changes in blood sugar throughout the day carefully. When there is large blood sugar, or it could spike after a meal, the patient needs to inject insulin into their bodies. When it is too low, fast-acting carbohydrate-rich food must be readily available.
To track this data, the patient must obtain blood samples through their fingers, manually log the results from their monitors, and inject insulin accordingly. Some people are privileged enough to access state-of-the-art tools such as continuous glucose monitors (CGMs), which can measure blood sugar through a small sensor placed under the skin. The readings generated by these devices are sent to pocket-sized monitors or smartphones. A steady stream of insulin is released from the pump clipped on a waistband. The CGM could interact with the insulin pump through a "close-loop" system, allowing it to adjust doses to maintain the blood sugar level within the target range.
Although these control algorithms work, they depend on hard-coded rules to make the device inflexible and reactive. Moreover, a CGM cannot send data if the patient forgets to bring the monitor.
Role of AI in Diabetes Management
A Ph.D. student at the University of Bristol's Department of Engineering Mathematics named Harry Emerson investigated the potential of machine learning in helping people with type 1 diabetes. He was diagnosed with type 1 diabetes before heading off to college. During that time, he needed to use medical devices to survive and go through the process of familiarizing himself with the technology.
Knowing the struggles encountered by diabetes patients in monitoring their blood sugar levels, Emerson collaborated with the University Hospital Southampton in teaching a machine learning algorithm that can save virtual patients. They used 30 simulated patients composed of 10 children, ten adolescents, and ten adults. These patients were synthesized by the UVA/Padova Type 1 Diabetes Simulator, a software approved by the FDA to replace preclinical testing in animals.
Their research team trained the AI on seven months of data and allowed it to learn the amount of insulin needed to deliver in various real-life situations. The team focused on reinforcement learning (RL) based on trial and error. After the offline training, they allowed the RL to decide the insulin dosing of the virtual patients.
According to Emerson, the offline RL successfully managed all the challenges in the simulator, even outperforming the current state-of-the-art controllers for diabetes patients. It also greatly improves when some needed data is missing or inaccurate.
RELATED ARTICLE : People With Type 1 Diabetes Are Six Times More Likely to Develop Dementia Later in Life
Check out more news and information on Diabetes in Science Times.
