The Challenge of Predicting Future Blood Glucose for Patients with Type I Diabetes

Abstract

Patients with Type I Diabetes (T1D) must take insulin injections to prevent the serious long term effects of hyperglycemia â high blood glucose (BG). These patients must also be careful not to inject too much insulin because this could induce hypoglycemia (low BG), which can be fatal. Patients therefore follow a âregimenâ that, based on various measures, determines how much insulin to inject at certain times. Current methods for managing this disease require the manual adjustment of a patientâs regimen over time based on the diseaseâs behavior (recorded in the patientâs diabetes diary). This is both time consuming and error-prone. If we can accurately predict a patientâs future BG values from their current features (e.g., predicting todayâs lunch BG value given todayâs diabetes diary entry for breakfast, including insulin injections), then it is relatively easy to produce an effective regimen. This study explores the challenges of BG modeling by applying a number of machine learning algorithms, as well as various data preprocessing variations (corresponding to 312 [learner, dataset] combinations), to a new T1D dataset that contains 30,221 entries from 51 different patients. Our most accurate predictor is a weighted ensemble of two Gaussian Process Regression (GPR) models where GPR#1 is learned using a patientâs entire history (over all meals) and GPR#2 is learned using data from individual meals. This ensemble achieved an errL1 loss of 2.72 mmol/L. This was an unexpectedly poor result given that one can obtain an errL1 of 2.94 mmol/L using the naive approach of only predicting the patients average BG. These results suggest that accurate BG prediction models may not be obtainable from the diabetes diary data that is typically collected; additional data may be necessary to build fine-grained BG control systems that use BG prediction models.