FutureG: Improving Predictability of Blood Glucose Levels in Type 1 Diabetics by using Lifestyle Data and
Machine Learning
Matthew Chang, Richa Laddha, Katherine Lee, Matt Lin, Kristine McLaughlin, Ria Shah, Jiming Song
Type 1 diabetes prevents the body from producing the insulin needed to break down glucose in the blood, requiring patients with type 1 diabetes to continually inject their own insulin. Because of this, blood glucose levels are often unpredictable and difficult to manage. Going outside of the normal Type 1 Diabetes blood glucose range of 70–250 mg/dL may cause complications such as heart disease, vision loss, and kidney disease. Maintaining an acceptable blood glucose level requires a patient to frequently monitor their blood glucose level and estimate the correct amount of insulin to give. Therefore, we are trying to predict glucose levels based on a user’s lifestyle and give feedback on how to maintain their ideal blood glucose range for a longer period of time. So far, to address our problem, we researched about factors that influence blood sugar and performed market research on current applications that help with blood glucose management. We found major factors include carbohydrates and bolus insulin (extra insulin given during meal times) and will be focusing on those as well as exercise in our model. We also found some applications that aim to predict blood glucose levels, but they have some limitations.
Market Research
OneDrop
OneDrop is an app that predicts if blood glucose levels will rise above 180 mg/dL or fall below 70 mg/dL using machine learning predictions. These predictions are based on data of people with similar health profiles and the app then gives advice about adjusting behavior. It is an accurate device, but it lacks the ability to make predictions that account for future meals or exercise.
Control-IQ
Control-IQ is a technology that algorithmically determines the correct amount of bolus insulin to inject using the trend of the blood glucose data. It provides a short-term, 30 minute prediction, but it also cannot account for future meals or activity. Another limitation of this technology is that it requires a specific insulin pump (t:slim X2) that our community partner does not have.
Our Solution:
We want to solve this problem by using machine learning to predict blood glucose levels for an individual with type 1 diabetes. In the morning, they would input the carbohydrate counts, times of their meals, times of exercise, and planned bolus insulin amounts for the day. Using this data, we would use their past blood glucose data to provide a predictive graph for the user for the rest of the day. This will allow the user to see when their blood glucose would likely spike or fall out of the ideal range and give them the opportunity to alter their actions and keep their blood glucose in a good range.
The user will be able to access this algorithm through a website. They can link their blood glucose data to the website, and it will have input boxes for future carbohydrate amounts and future insulin amounts.
Design Review Feedback and Future Goals:
Algorithm:
Our community partner recommended creating a script to evaluate the user’s insulin dose after every meal before designing our predictive machine learning model. The goal of the script is to determine whether the user gave themselves the correct amount of insulin to stay within an ideal blood glucose range of 70 to 170 mg/dL and find a mathematical correlation between the user’s bolus injection amount, their carbohydrate intake, and the effect on their blood glucose level.
Our immediate goal is to design and code the script to assess if the amount of insulin taken at a certain time period was correct, and if not, then how much insulin the user should have taken. We hope to have a version of this script done by Design Review 2.
A second goal is to gain access to larger datasets and to identify which machine-learning algorithms would be most appropriate for the task. Then we can produce a website that can give a predictive graph for the day.
Web Design:
Simplify the website design by reducing the format to a single page and creating a two column design.
For the website, one of our short-term goals is to determine what language we want to create the website with. We also want to determine how to link the algorithm to the website and create the desired graphs. We will test these concepts initially using fake data.
