Enhancing Type 1 Diabetes Management Through Machine Learning

Type 1 diabetes (T1D) is an autoimmune condition where carbohydrate metabolism is disrupted due to insufficient insulin secretion. In the pursuit of metabolism, it is important to maintain tight glycemic control. Current insulin pumps require user input for dose adjustments, leading to suboptimal glycemic control and artificial pancreas are expensive and fail to account for various factors. This study investigates the application of machine learning (ML) for insulin delivery in T1D patients. The purpose is to support underserved T1D patients with an ML control solution for better diabetic management and better health. We propose regression models utilizing time to predict insulin dosages based on a large patient dataset. To apply to patients across the world an app was built that implements this algorithm with insulin dosage suggestions that continuously learn through a feedback loop. The optimal regression model was an LSTM model trained on data from only one patient. The model achieved a root mean squared error of 1.0055. In addition to the model, a feedback loop was implemented for the model to adapt to the specific patient who is using the app for insulin dosage suggestions. This has yet to be tested out in a study.