A totally data-driven whole-brain multimodal pipeline for the discrimination of Parkinson's disease, multiple system atrophy and healthy control

Abstract

Parkinson's Disease (PD) and Multiple System Atrophy (MSA) are two parkinsonian syndromes that share many symptoms, albeit having very different prognosis. Although previous studies have proposed multimodal MRI protocols combined with multivariate analysis to discriminate between these two populations and healthy controls, studies combining all MRI indexes relevant for these disorders (i.e. grey matter volume, fractional anisotropy, mean diffusivity, iron deposition, brain activity at rest and brain connectivity) with a completely data-driven voxelwise analysis for discrimination are still lacking. In this study, we used such a complete MRI protocol and adapted a fully-data driven analysis pipeline to discriminate between these populations and a healthy controls (HC) group. The pipeline combined several feature selection and reduction steps to obtain interpretable models with a low number of discriminant features that can shed light onto the brain pathology of PD and MSA. Using this pipeline, we could discriminate between PD and HC (best accuracy = 0.78), MSA and HC (best accuracy = 0.94) and PD and MSA (best accuracy = 0.88). Moreover, we showed that indexes derived from resting-state fMRI alone could discriminate between PD and HC, while mean diffusivity in the cerebellum and the putamen alone could discriminate between MSA and HC. On the other hand, a more diverse set of indexes derived by multiple modalities was needed to discriminate between the two disorders. We showed that our pipeline was able to discriminate between distinct pathological populations while delivering sparse model that could be used to better understand the neural underpinning of the pathologies.

Keywords: Data-driven clinical classification; Multimodal MRI; Parkinsonism discrimination.

Fig. 1

Predictive pipeline. Colored brains represent the different indexes used for prediction, brains of the same color represent indexes from the same MRI modality. The outer malva square represents the outer 10-folds CV scheme while the inner orange square represents the inner 10-folds CV set up to find the best combination of modalities. Green squares represent features selection and reduction steps while grey squares represent preprocessing, model fitting and prediction steps, grey ovals represent the intermediate and final outcome of the pipeline.

Fig. 2

Comparison of the performance of the different discrimination tasks and cluster extent.

Fig. 3

Frequency of occurrence of the modalities and their combination in 100 folds (10 folds CV repeated 10 times) for the discrimination between PD and HC. globalCorr = global correlation; localCorr = local correlation; alff = fraction of alpha low frequency fluctuations, fa = fractional anisotropy, gm = grey matter volume, md = mean diffusivity.

Fig. 4

Most frequently selected voxels for the most frequently selected modalities (PD vs HC). A) fALFF; B) global correlation; C) local correlation.

Fig. 5

Frequency of occurrence of the modalities and their combination in 100 folds (10 folds CV repeated 10 times) for the discrimination between MSA and HC. globalCorr = global correlation; localCorr = local correlation; alff = fraction of alpha low frequency fluctuations, fa = fractional anisotropy, gm = grey matter volume, md = mean diffusivity, r2s = R2*.

Fig. 6

Most frequently selected voxels for the most frequently selected modalities (MSA vs HC). A) MD; B) r2s.

Fig. 7

Frequency of occurrence of the modalities and their combination in 100 folds (10 folds CV repeated 10 times) for the discrimination between PD and MSA. globalCorr = global correlation; localCorr = local correlation; alff = fraction of alpha low frequency fluctuations, fa = fractional anisotropy, gm = grey matter volume, md = mean diffusivity, r2s = R2*.

Fig. 8

Most frequently selected voxels for the most frequently selected modalities (PD vs MSA). A) gm; B) FA; C) global correlation; D) MD.

Fig. 9

reports the performance of the best model together with its cluster extent for each discrimination task (upper panel). In the middle panel are reported the modalities most frequently selected for each discrimination task (the brain slices are from a representative subject and intensity coded). In the lower panel are reported the cluster most frequently observed (> 50 folds, excepts for R2* > 40 folds) for each of the most observed modalities. Spatial cluster for global correlation for the discrimination between PD and MSA are not shown as no voxel was observed in >25 folds.