Data completeness and transformer-based imputation for OpenAPS diabetes data.
Previous work on OpenAPS data produced a suite of generalised linear models and sophisticated machine learning models for blood glucose prediction. OpenAPS was developed by the #wearenotwaiting community to provide a flexible and customisable approach to diabetes management.
The data is not always complete, which means models are often trained on limited subsets. This project investigated missingness in key variables and then evaluated statistical and machine learning imputation approaches, with progression toward transformer-based imputation methods.