Optimizing Glycemic Control in Type 1 Diabetic Patients using a Deep Learning-Based Artificial Pancreas with a Secure Glucagon and Insulin Delivery System

Abstract

Type 1 diabetes impacts millions worldwide, with some patients facing rapid fluctuations in their blood sugar levels. These fluctuations can negatively impact an individuals quality of life and if untreated, can lead to nerve damage, coma, and death. While current methods have helped address hyperglycemia (high blood sugar), there has been less success with hypoglycemia (low blood sugar) and glucagon administration. To bridge this gap, an artificial pancreas with a novel insulin and glucagon pump was developed. Initiating the system, a personalized mobile app enables users to input meal carbohydrate and insulin bolus data. The data is then transmitted to a deep learning model that incorporates Continuous Glucose Monitor readings, along with carbohydrate and insulin data from the app. The two-layer Long Short-Term Memory network, developed in Python, accurately forecasts blood sugar levels on Ohio Universitys OhioT1DM patient dataset and the UVa/Padova simulations data for a 30 minute-interval. An algorithm then utilizes the predictions to calculate optimal insulin and glucagon doses using metabolization formulas. To ensure system security, data is transmitted through a cloud-based MQ Telemetry Transport server and secured with industry-standard authentication and encryption methods. Finally, a microcontroller-based prototype accurately dispenses insulin and glucagon doses. The system kept in-silico patients at optimal levels for 38% longer and reduced dangerous levels by 22% compared to conventional controllers on an FDA-approved preclinical trial alternative simulation. By addressing both hypo and hyperglycemia, this real-time medical device can be a transformative tool for individuals with diabetes, enabling them to live healthier and more fulfilling lives.