DTI datasets of 162 PD patients (age: 63.9 ± 9.3 years; gender: 34.2% female; disease duration: 6.5 ± 4.1 months; mean MDS-UPDRS-III: 13.9 ± 2.1; mean Hoehn and Yahr stadium: 1.2 ± 0.3) and 70 age and gender-matched healthy controls (HC) (age: 62.1 ± 10.1 years; gender: 34.9% female) were analyzed. This study used human subject recordings chosen from the Parkinson’s Progression Marker Initiative (PPMI) database. The PPMI dataset was published open-access with a positive ethics statement of the responsible authorities. Therefore, additional ethics committee approvals do not apply to this study. DTI-MR sequences were acquired on a Siemens 3 T TIM Trio scanner using a 12-channel matrix head coil and a two-dimensional echo-planar DTI sequence (TR/TE = 900/88 ms, flip angle = 90°, voxel size = 2 × 2 × 2 mm
3, 72 slices, 64 gradient directions with a b-value of 1000 s/mm
2). In addition, a non-gradient volume (b = 0 s/mm
2) was acquired as well. Further details of the PPMI image acquisition protocol can be seen online (
http://www.ppmi-info.org/wp-content/uploads/2017/06/PPMI-MRI-Operations-Manual-V7.pdf). We performed pre-processing by using the PANDA-toolbox (v1.3.1) in Matlab 2018b, including normalization to standard space (via FMRIB58_FA template, 2 mm × 2 mm × 2 mm voxel size) [
6]. In addition to conventional diffusion metrics (FA, MD, AD, and RD), we calculated local diffusion homogeneity (LDH) as another measure of microstructural white matter integrity. For the interpretation of DTI images, we calculated the following standard diffusion metrics based on the three-dimensional diffusion of water as a function of spatial location: Fractional Anisotropy (FA) is a summary measure for interpreting microstructural integrity. Mean Diffusivity (MD) is a measure of the cell membrane density. It is, therefore, sensitive for cellularity, edema, and necrosis of investigated tissue. Axial Diffusivity (AD) decreases in axonal injury. Radial Diffusivity (RD) increases in de- or dysmyelination of axons. A concise review article on the interpretability of diffusion metrics to investigate microstructural grey and white matter changes are described in a review article by Alexander et al. [
1]. Local diffusion homogeneity (LDH) is another diffusion metric that is specifically relevant to assess tissue homogeneity based on neighboring voxels [
9]. We computed LDH for 6, 18, and 26 neighboring voxels using Spearman’s Rank Correlation coefficient (06LDHs, 18LDHs, and 26LDHs) and Kendall’s coefficient concordance (06LDHk, 18LDHk, and 26LDHk) [
9]. Voxel-wise whole-brain analysis was performed using the FM-RIB58_FA template. We performed ROI-labeled analyses based on the well-established AAL atlas [
21]. To further increase the signal-to-noise ratio, we additionally performed classification after masking of the SN using the ATAG atlas for the elderly population [
10]. The datasets were classified through bSVMs (for single modalities) as well as MKL (for concatenated modalities). Ten-fold cross-validation (CV) and nested (leave one subject out) hyperparameter optimization as implemented in the PRoNTo-Toolbox (v2.1) [
18]. The determination of relevant bSVM and MKL parameters (such as the applied L1 regularization method or the nested hyperparameter optimization) is following standard practice and is extensively described in the publications of Schrouff et al. [
17,
18]. Age, gender, and total intracranial volume were used as covariates. Balanced Accuracy (BA) and area under the curve of the receiver-operating characteristic curve (ROC-AUC) were calculated to assess classification performance and were compared to random permutation testing (against 10.000 permutations).