Hey HN!

Currently, the key parameters for Type 1 diabetes management, like Insulin-To-Carb ratio (ICR) and Correction Factor (CF), are estimated manually by eyeballing historical charts and making a subjective guess about what the "true" values should be.

There is a lot of uncertainty and noise in the data itself, so the estimation task is hard even for Diabetic doctors.

I thought there had to be a more principled way to make such decisions from the data, so I built a Bayesian ML model that models data the same way these parameters are supposed to be used (i.e., a linear model assumption). Essentially, it tries to answer the question: "What are the best ICR and CF values to use given the observed noisy data?"

You can try the model on the test data in Google Colab. See the README for details.

I'm looking for feedback and collaboration to validate the model on different patient data.