Multi-modal Biomarkers

These projects attempt combining evidence from multiple sources to improve classification performance.

AddNeuroMed

A dataset containing different biomarker modalities (gene expression, genotype and structural MRI data) from participants with Alzheimer's type dementia, mild-cognitive impairment and others who are cognitively normal.

The challenges here are:

• How to combine modalities when a given participant has only one or two of the biomarkers, but not necessarily the full complement ?
• Is it better to use 'brute force' combination (train a giant classifier, let it sort out combinations of features across modalities) or to train 'expert classifiers' on each modality, and combine their predictions ?

A preliminary report on the problem, approach and methods can be found here, and a presentation (intended for a clinical audience) here.

MUTRIPS

I help out with some model selection/validation work (mainly for classifiers) for the project team.  MUTRIPS investigates the biology underpinning treatment resistance in psychotic illness.  My interests are how to use machine learning (predictive) methods that are transparent (can we be sure what the models are learning) and ultimately, robust as clinical decision support tools (i.e. for individual-person level prediction, rather than using classical inference and merely claiming predictive utility). 

Some preliminary work using results from  colleagues' analyses of magnetic resonance spectroscopy and structural imaging (using regularised GLMs) are here, and using trees and boosting here (as R notebooks / work in progress - they are fairly tutorial).