Feature selection/visualisation of ADNI data with iterative partial least squares

This article introduces a variable selection and visualisation approach for medical imaging big data analysis based on Partial Least Squares, dubbed Picky Partial Least Squares. The method can handle very high-dimensional data and appears to be able to find relevant clusters of data points. It has been developed to deal in particular with large datasets. The method is validated experimentally on medical images from the ADNI (Alzheimer´s Disease Neuroimaging Initiative). It is shown to perform better than standard PLS on the datasets and identifies relevant brain areas and SNPs as linked to Alzheimer´s Disease. In particular the temporal lobes of the brain are highlighted by the algorithm, along with SNPs such as rs157580, which have previously been linked to Alzheimer´s Disease. The method is also able to classify Alzheimer´s patients from controls directly from the original high-dimensional imaging data, without any feature selection and dimension reduction. Unlike existing publications, the focus of this paper will be to select and visualise the image features that PPLS considers as related to Alzheimer´s Disease.