Parametric Bayesian multi-task learning for modeling biomarker trajectories

This model builds and tests longitudinal trajectory models for multiple subjects at once, allowing for information sharing (i.e. coupling) of subjects' models using biomarker similarity measures. This code is from our "Modeling longitudinal biomarkers with parameteric Bayesian multi-task learning" (paper) and OHBM 2018 poster.

Directories

The blr_sim directory contains the top-level files used for simulations, while the blr directory contains most of the model training, prediction and performance assessment code. The gpml-matlab-v4.0-2016-10-19 directory from the gpml toolbox is used for hyperparameter optimization, the aboxplot directory from Alex Bikfalvi is used for making nice boxplots. utils contains some basic utility functions.

Simple example

Within the blr_sim directory is a simple example that you can run and modify:

simple_example

Simulations

You can also run the simulations described in our paper via:

sim_both_full

This will generate a couple of intermediate files for you in the out_blr_sim directory along with two figures (from the paper) that show mean absolute error (MAE) and parameter inference related metrics across the 50 simulation runs and 2 simulation scenarios (intercept-variation and slope-variation in subjects' trajectories).

This above command will take at least a few hours to run as it's building eight coupled models fifty times for four different noise levels, across two simulation scenarios (8 x 50 x 4 x 2 = 3,200 models for 200 subjects).

You can run a faster version that will give you a good idea of how it all works by running:

sim_both_quick

This will build eight models ten times for the four noise levels and two scenarios for fewer subjects (8 x 10 x 4 x 2 = 640 for 100 subjects). It should run in minutes rather than hours.

If you want to replicate our ADNI results, you need to have access to the ADNI dataset. Once you do, I can send you the necessary data and code. Contact: l.aksman@ucl.ac.uk.