Lead Investigator: Garrett Ash, Yale School of Medicine
Title of Proposal Research: Developing a pre-trainable foundation time series model for blood glucose prediction among people with type 1 diabetes (T1D) and then applying it to a vulnerable, understudied subset of them
Vivli Data Request: 9848
Funding Source: The lead researcher will be supported by the National Institute of Diabetes, Digestive, and Kidney Diseases of the National Institutes of Health under a mentored research scientist development award (K01DK129441).
Potential Conflicts of Interest: None
Summary of the Proposed Research:
Type 1 diabetes (T1D) is an autoimmune condition held by ~8 million people where the pancreas does not make its own insulin, meaning insulin must be delivered exogenously creating a need to constantly balance insulin, diet, and exercise to maintain blood glucose control. Prediction of upcoming trends in blood glucose is the cornerstone of the burgeoning field of automated insulin delivery devices and glucose management decision support systems for T1D. This prediction is made using current and past blood glucose and insulin-on-board and can be further improved with additional inputs including diet logs and smartwatch readings (steps, exercise, heart rate, heart rate variability, skin temperature). Several dozen studies have developed and refined glucose prediction algorithms with approaches that include physiological models, traditional machine-learning models, deep learning models, and hybrid physiological + machine-learning models (reviewed in (Ahmed et al., 2024)). The datasets to train and test these models (i.e., people with T1D wearing trackers under free-living conditions) have grown in recent years (Ahmed et al., 2024) but are limited by underrepresentation of the subgroups of people with T1D most in need of clinical support, specifically those with chronically elevated blood glucose levels, non-White race/ethnicity, and/or adolescent age. Even the latest dataset that is best suited to address these limitations (T1DEXIP, n=250 adolescents (Riddell et al., 2024)) has just 49% of individuals with average glucose above recommended levels (HbA1c >7.0%), 17% moderately above recommended levels (HbA1c >8.0%), and 1% very above recommended levels (HbA1c >9.0%). Nationally, these proportions are much higher: 94%, 72%, and 44% (Foster et al., 2019).
Our research group has historically pioneered interventions providing exercise support to adolescents with T1D, elevated HbA1c, low socioeconomic status, and diverse race/ethnicity (Ash et al., 2019; Ash et al., 2024; DeJonckheere et al., 2021). In the most recent of these studies, participants wore trackers for continuous glucose monitoring, insulin dosing, physical activity, heart rate, and sleep. However, this work is in early feasibility phase with sample sizes n=10-18 participants and recruiting larger samples can be quite expensive; in particular, these families have high opportunity cost to participating in research because of competing demands.
It would be beneficial if machine-learning could leverage the large datasets for foundation models that could then be applied to smaller datasets of individuals in greatest need. The pre-trained foundation time series models are suited to this purpose by building foundation models that can then be applied to smaller datasets. Examples are TimeGPT-1 (Garza & Mergenthaler-Canseco, 2023), prompt-based generative pre-trained transformer for time series forecasting (TEMPO) (Cao et al., 2023), and time series forecasting by reprogramming large language models (Time-LLM) (Jin et al., 2023). We therefore plan the following sequence of steps. First, we will train a foundation model to predict glucose using multimodal inputs (past and current glucose, carbohydrates, insulin, steps, heart rate, sleep) from large datasets (T1DEXI (Riddell et al., 2023), T1DEXIP (Riddell et al., 2024)). Second, we will apply the foundation model to a dataset of our participants who have diversity of race/ethnicity and glucose control (Home-Based Virtual Activity Program for Youth with Type 1 Diabetes; HAP-V-T1D). In both applications, we will test the added value of inputs that are always collected in routine T1D care (glucose, insulin dosing) versus those that require extra effort by the participant (fitness smartwatch metrics, diet logs) to determine the priority of encouraging individuals to wear these devices.
The ultimate goal of this research is a modeling framework that leverages existing datasets to create a more equitable approach to predicting blood glucose and therefore to developing glucose stabilizing technology for T1D. As well, the approach could be taken to other diseases where the large databanks have sampling bias.
Requested Studies:
Type 1 Diabetes EXercise Initiative: The Effect of Exercise on Glycemic Control in Type 1 Diabetes Study
Data Contributor: Jaeb Center for Health Research Foundation, Inc.
Study ID: T1-DEXI
Sponsor ID: T1-DEXI
Type 1 Diabetes EXercise Initiative Pediatric Study (T1DexiP): The Effect of Exercise on Glycemic Control in Youth with Type 1 Diabetes
Data Contributor: Jaeb Center for Health Research Foundation, Inc.
Study ID: T1-DEXIP
Sponsor ID: T1-DEXIP