Uniform Client Selction Over Time Series Medical Task

[rtaiello]

With @luciainnocenti, we completed this tutorial, where we evaluated Fed-Biomed using uniform client selection using a medical time series dataset over a large number of nodes.

The dataset used is the Replace-BG dataset. More information can be found here:

• Glucose Prediction using Replace-BG Dataset

The preprocessing is inspired by the following paper:

• Long-Term Prediction of Blood Glucose Levels in Type 1 Diabetes Using a CNNLSTM-Based Deep Neural Network

The code is inspired by the following GitHub repository:

• GluPred GitHub Repository

In this tutorial, we can add up to 202 nodes and start up to 180 nodes (uniform selection of 10%) for federated training. We then evaluate the global model over 22 clients. The training runs for 400 rounds, with the following requirements: RAM > 80 GB and 64 cores. This setup takes around 10 hours to complete.
Here the result:

However, in the notebook, it is possible to decrease the number of started nodes, we will use 10 nodes for training (uniform selection of 50%), and for testing for 5 rounds.