Development and Evaluation of Personalized Explainable Machine Learning Models to Predict and Prevent Nocturnal Hypoglycemia in Type 1 Diabetes

点击翻译按钮获取中文摘要

Development and Evaluation of Personalized Explainable Machine Learning Models to Predict and Prevent Nocturnal Hypoglycemia in Type 1 Diabetes Project Summary Hypoglycemia (glucose < 70 mg/dL) remains the limiting factor for achieving optimal glycemic control in type 1 diabetes (T1D), with nocturnal hypoglycemia being particularly dangerous. Nocturnal hypoglycemia may result in physical injury, poor sleep quality, fear of hypoglycemia, and hypoglycemia unawareness. Severe episodes can cause seizures and unconsciousness requiring emergency care, and even death (dead in bed syndrome). While automated insulin delivery (AID) systems have shown benefits in glucose control during the night, nighttime hypoglycemia still occurs. Moreover, many people with T1D manage their glucose with continuous subcutaneous infusion pump (CSII) therapy or multiple daily insulin injections (MDI) therapy. Data updated between 2013 and 2014 from 16,061 individuals with T1D participating in the T1D Exchange clinic registry showed that approximately 40% participants managed their glucose with MDI. In this project, we propose to develop and evaluate a personalized decision support tool that collects and analyzes glucose measurements, insulin, meals, and physical activity data to predict at bedtime the likelihood of overnight hypoglycemia and recommend a proactive carbohydrate intervention to substantially reduce nocturnal hypoglycemia. In the engineering development phase of the project, we will use unique datasets of time-matched glucose management data (i.e., continuous glucose measurements, insulin, meals, and exercise) from pump, closed- loop and MDI users to extract information about the major contributors to nocturnal hypoglycemia risk and train a population-based prediction model that will be personalized over time to better capture inter-subject variability. We will design a bedtime intervention consisting of a bedtime smart snack with variable nutrient content that can prevent nighttime hypoglycemia. Snacks will vary by macronutrient content and size to optimize time to peak post-prandial glycemia that will match the timing to predicted episode of hypoglycemia. We will conduct a randomized cross-over study to evaluate our smartphone-based decision support tool on a cohort of 20 people with T1D who are MDI users and are at higher risk of experiencing hypoglycemia. Participants will be randomly assigned to either first use CGM only (control period) followed by a smartphone- based decision support tool + nocturnal hypoglycemia intervention (intervention period), or vice-versa. The control and intervention periods will have a duration of three weeks each. We will measure the effect of the intervention by comparing the percent time in nocturnal hypoglycemia during the control period vs. the intervention period. We will also retrospectively measure the accuracy of the prediction model in predicting nocturnal hypoglycemia using data from the control period. We expect that the proposed bedtime intervention will lead to a significant reduction in time spent in hypoglycemia overnight of at least 50% reduction relative to baseline.