Model Training

!!! warning "Under Construction" This documentation is currently under active development and subject to change. Some sections may be incomplete or missing.

Training Models

For most of our modeling tasks, the training falls into two broad categories, the models training goal, either single patient or multi-patient, and the model type, which can be either statistical, ml, dl, or foundation based models.

Training Goal

Single Patient Training

Train one model for each patient. The goal is to create a model that minimizes its error as much as possible for a single patient. Often it makes sense to take an existing multi-patient model and fine-tune it for a specific patient.

Multi-Patient Training

Train one model for all patients. The goal is to create a generic model that works as well as possible for all patients.

Model Type

Statistical Modeling

Currently, all of our statistical modeling is done with sktime. Essentially, these mostly fall under classic statistical time series techniques, almost anything that is not deep learning or a foundation model falls under this category, e.g.,: naive, ARIMA, exponential, etc We want to exhaustively find the best modeling parameters for each family of models.

Machine Learning

Models that aren't statistical modeling or deep learning, for e.g.,: Gradient Boosting, Tree-based algorithms, etc.

Deep Learning

Currently, all of our deep learning modeling is done with sktime wrappers, but this will be expanding to other time series libraries, usually within the sktime ecosystem like pytorch forecasting. Excluding foundation models, all neural network model training belongs here.

Foundational

There are two main foundation model use cases with our projects: 1. Fine-tuning existing foundation models with diabetes data - We plan to try LagLLama, Moirai, Chronos, TimesFM, TimeGPT, TimeGPT-Long, TinyTimeMixer, and potentially others. 2. Training diabetes foundation models from scratch with as much diabetes data that we can find.