Diabetes is a chronic condition that affects millions of people throughout the world and studies show that there is a gap between research and smartphone-based solutions. This thesis develops a novel model, ICARUS, using Machine Learning techniques to predict blood glucose levels in Type 1 Diabetes Mellitus (T1DM) people, designed for smartphones and sensor-based devices such as wearables. Through a systematic literature analysis to evaluate what parameters affect blood glucose levels previous medical studies were included and existing algorithms were investigated. ICARUS model was developed by using an ensemble method to combine two Machine Learning techniques (RandomForests and LSTM) achieving adaptability and low prediction error. Designed with mobile constraints in mind, such as battery and memory usage, ICARUS represents a paradigm shift in diabetes care, offering a lightweight, offline, and accurate model for individuals with limited internet access. Rigorously evaluated under real-life-based scenarios, including missing data instances, individual patient training, and missing features, ICARUS demonstrated efficient performance with an average RMSE value of 16.479, outperforming similar studies. Consequently, this thesis delivers a model that serves as an initial step for future studies in different areas, expanding the current diabetes-based scenario into a broader population.
- machine learning
- artificial intelligence
- ensemble
- Long Short-Term Memory
- RandomForests
- Deep Learning
- diabetes
- blood glucose prediction
- digital health
- prediction model
- sensors
- mobile
Blood Glucose Prediction Model for Smartphones and Sensor-Based Devices
L G Costa, W. (Author). 20 Jun 2024
Student thesis: Doctoral Thesis