Evaluation of Machine Learning Algorithms and Structural Features for Optimal MRI- Based Diagnostic Prediction in Psychosis

A relatively large number of studies have investigated the power of structural magnetic resonance imaging (sMRI) data to discriminate patients with schizophrenia from healthy controls. However, very few of them have also included patients with bipolar
disorder, allowing the clinically relevant discrimination between both psychotic diagnostics. To assess the efficacy of sMRI data for diagnostic prediction in psychosis we objectively evaluated the discriminative power of a wide range of commonly used machine learning algorithms (ridge, lasso, elastic net and L0 norm regularized logistic regressions, a support vector classifier, regularized discriminant analysis, random forests and a Gaussian process classifier) on main sMRI features including grey and white matter voxel-based morphometry (VBM), vertex-based cortical thickness and volume, region of interest volumetric measures and wavelet-based morphometry (WBM) maps. All possible combinations of algorithms and data features were considered in pairwise classifications of matched samples of healthy controls (N = 127), patients with schizophrenia (N = 128) and patients with bipolar disorder (N = 128). Results show that the selection of feature type is important, with grey matter VBM (without data reduction) delivering the best diagnostic prediction rates (averaging over classifiers: schizophrenia vs. healthy 75%, bipolar disorder vs. healthy 63% and schizophrenia vs. bipolar disorder 62%) whereas algorithms usually yielded very similar results. Indeed, those grey matter VBM accuracy rates were not even improved by combining all feature types in a single prediction model. Further multi-class classifications considering the three groups simultaneously made evident a lack of predictive power for the bipolar group, probably due to its intermediate anatomical features, located between those observed in healthy controls and those found in
patients with schizophrenia. Finally, we provide MRIPredict (http://www.mripredict.com/) a free tool for SPM, FSL and R, to easily carry out voxelwise predictions based on VBM images.

The following directories are included in the compresed files:

/rois: It contains a single text file (rois1.txt) with volumes of cortical and subcortical structures for each one of the subjects.

/vbm_gm: It contains the grey matter probability maps for the three groups of subjects in format NIFTI compressed (nii.gz extension).

/vbm_wm: It contains the white matter probability maps for the three groups of subjects in format NIFTI compressed (nii.gz extension).

/thickness: It contains two binary files (thickness_lh.bin and thickness_rh.bin) which contain cortical thickness data for the left hemsiphere (lh) and right hemisphere (rh) for all subjects included in the study.

Specifically, cortical thickness data is stored in a 4 byte real format (149955 points per subject (lh), and 149926 points per subject (rh), with the 383 subjects stored sequentially in the same order as in the other data formats.  

/volume: Two files containing cortical volume information with the same format as for thickness (volume_lh.bin and volume_rh.bin).