Show HN: I have trained 15 blood glucose forecasting models on my own data (github.com)

A recent project showcased on Hacker News highlights the development of 15 blood glucose (BG) forecasting models, including a self-designed transformer model, by an individual managing type 1 diabetes. These models aim to predict blood glucose levels two hours in advance, utilizing both synthetic data and publicly available datasets like MetaboNet. The main model, referred to as T1DMAI D16 DILATE, incorporates innovative training techniques such as LoRA (Low-Rank Adaptation) to enhance performance, achieving competitive results against established models like XGBoost. This endeavor is significant for the AI and ML community as it demonstrates the potential of personalized health management through machine learning. By leveraging synthetic continuous glucose monitoring (CGM) traces, the models offer insights into the immediate challenges faced by individuals with diabetes, pushing the boundaries of traditional forecasting methods. The performance metrics reveal that although XGBoost remains the top performer, the D16 DILATE model closely follows, particularly excelling in 90-minute forecasts with the lowest root mean square error (RMSE). The work emphasizes the importance of fine-tuning models to adapt to individual data and highlights the practical application of AI in medical contexts, albeit with the caution that these forecasts are not clinically validated.