The goal of this project is to develop machine learning models that leverage underutilized data and account for individual patient differences to improve diabetes management. Diabetes is a chronic disease, which must be treated and managed over a lifetime. In type 1 diabetes, the pancreas does not produce insulin, an essential hormone needed to convert food into energy. The disease is treated with insulin and managed by monitoring and controlling blood glucose levels. Good blood glucose control is key to avoiding serious diabetic complications, but achieving and maintaining good blood glucose control is difficult. Patients are highly individual in their responses to treatment and to life events that affect blood glucose levels. Large volumes of blood glucose data are collected automatically, but automated analysis is lacking. Patients do not always know when problems are impending; problems occurring while patients are asleep are especially dangerous. Machine learning models that predict blood glucose levels would enable or facilitate new applications of direct benefit to patients, including: alerts to immediately notify patients of impending problems; decision support systems recommending actions to prevent problems; and educational simulations showing the effects of different treatment choices or lifestyle options on blood glucose levels.
The task of blood glucose prediction is approached as a time series forecasting problem. Blood glucose is predicted based on a patient's prior blood glucose levels, insulin data, meal data, exercise data, sleep patterns and work schedules. Batch and incremental time series regression models, including support vector machines and neural networks, are being investigated. To account for individual patient differences, separate models are trained for each patient. However, transfer learning may enable data from multiple patients to aid in building models for patients with limited historical data. Models are sought that are robust in the face of imperfect data, including missing life events, inaccurately recorded life events, and noisy glucose sensors. Disjoint sets of training and testing data extracted from non-overlapping time intervals are used to build models that are compared against baselines such as autoregressive integrated moving average models. Standard metrics for comparing models, such as root mean square error and coefficient of determination, will be supplemented with domain dependent measures of goodness, including the Clarke error grid.
This work aims to improve the overall health and wellbeing of the nearly two million Americans with type 1 diabetes. Predicting blood glucose problems in advance gives patients time to take steps to prevent the predicted problems from occurring. This improves blood glucose control, which is known to reduce the risk of serious diabetic complications, including blindness, amputations, kidney disease, strokes, heart attacks, and death from severe hypoglycemia. In addition, this research project forms the cornerstone of a new Smart Health and Wellbeing Laboratory at Ohio University. This new laboratory is designed to promote further interdisciplinary research among computer scientists and health care professionals as well as to attract more women to careers in computer science. Additional information about the project and the laboratory is available at http://oucsace.cs.ohiou.edu/~marling/shb.html.