Publication Date
Fall 2024
Degree Type
Master's Project
Degree Name
Master of Science in Computer Science (MSCS)
Department
Computer Science
First Advisor
Sayma Akther
Second Advisor
Robert Chun
Third Advisor
Fabio Di Troia
Keywords
Diabetes Prediction, Time-Series Exploration, deep learning
Abstract
Diabetes is a lifelong illness that, if not detected or managed appropriately, turns into serious complications. Correct glucose forecasting is critical to ensuring timely interventions, thereby minimizing risks of hyperglycemia and hypoglycemia, and optimizing the management strategies of the disease. Classical machine learning models have been applied in the blood glucose forecasting problem for a long time, however, usage of transformer-based architectures is still scarce within the literature. Due to the self-attention mechanism, transformers can capture temporal relationships very effectively, which makes them suitable for time-series data. TFT is a novel framework proposed here to utilize time-series data from CGM systems. TFT utilizes self-attention to analyze complex time and feature correlations and to learn both short and long-term dependency patterns and applies multi-level attention to both temporal and static covariates to enhance interpretability and model performance. TFT also integrates exogenous variables (such as meals, and insulin) which is useful inBG prediction. Using a transfer learning approach, we pre-train the TFT on the OhioT1DM dataset and then fine-tune it on the D1NAMO dataset. We further perform the comparative experiments with LSTM, GRU and DRNN deep learning models. The model architecture allows us to predict BG levels at 15 and 30-minute prediction horizons. Key metrics like root mean square error and mean absolute error are taken to evaluate the models. This work progresses BG times series forecasting by exploring Temporal fusion transformer and deep learning models under a multi-horizon forecasting setup.
Recommended Citation
Meduri, Monica, "Temporal Dynamics in Diabetes Prediction: A Sensor-Driven Time-Series Exploration" (2024). Master's Projects. 1429.
https://scholarworks.sjsu.edu/etd_projects/1429
