Show HN: I have trained 15 blood glucose forecasting models on my own data

This project is a research “model zoo” comparing blood-glucose forecasting approaches, built by a person with type 1 diabetes to explore whether models could support day-to-day management. It evaluates classical methods, neural networks, and pretrained forecasting models on the developer’s own continuous glucose monitor data and the public MetaboNet dataset. The developer’s model, T1DMAI, is a small 31,000-parameter transformer that predicts glucose up to two hours ahead and uses glucose, carbohydrate, and insulin signals. It is first trained on simulated data, then optionally adapted to real data with a 3,771-parameter LoRA adapter. The benchmark also examines prediction intervals, early warning time, and detection of low glucose, not just average forecast error.

The main finding is that no model performs well across all settings, and XGBoost is generally strongest. On the developer’s test data, T1DMAI with LoRA ranks second among 22 methods and is statistically comparable to XGBoost at 60- and 90-minute horizons; on MetaboNet it ranks seventh, though it has the lowest RMSE at 90 minutes. Fine-tuning improves 60-minute scores on both datasets, but can worsen longer-horizon results. T1DMAI gives earlier warnings and catches more lows than XGBoost on the developer’s data, while nearly all low-RMSE methods miss lows in MetaboNet. Results are limited by small personal-data coverage, differing dataset conditions, and access to future insulin and carbohydrate inputs in the personal-data evaluation. The software is explicitly for research and education, not clinically validated or suitable for treatment decisions.