Multiclass characterization of frontotemporal dementia variants via multimodal brain network computational inference

Abstract Characterizing a particular neurodegenerative condition against others possible diseases remains a challenge along clinical, biomarker, and neuroscientific levels. This is the particular case of frontotemporal dementia (FTD) variants, where their specific characterization requires high levels of expertise and multidisciplinary teams to subtly distinguish among similar physiopathological processes. Here, we used a computational approach of multimodal brain networks to address simultaneous multiclass classification of 298 subjects (one group against all others), including five FTD variants: behavioral variant FTD, corticobasal syndrome, nonfluent variant primary progressive aphasia, progressive supranuclear palsy, and semantic variant primary progressive aphasia, with healthy controls. Fourteen machine learning classifiers were trained with functional and structural connectivity metrics calculated through different methods. Due to the large number of variables, dimensionality was reduced, employing statistical comparisons and progressive elimination to assess feature stability under nested cross-validation. The machine learning performance was measured through the area under the receiver operating characteristic curves, reaching 0.81 on average, with a standard deviation of 0.09. Furthermore, the contributions of demographic and cognitive data were also assessed via multifeatured classifiers. An accurate simultaneous multiclass classification of each FTD variant against other variants and controls was obtained based on the selection of an optimum set of features. The classifiers incorporating the brain’s network and cognitive assessment increased performance metrics. Multimodal classifiers evidenced specific variants’ compromise, across modalities and methods through feature importance analysis. If replicated and validated, this approach may help to support clinical decision tools aimed to detect specific affectations in the context of overlapping diseases.


Supplementary data Atrophy patterns
The SBM analysis was used with the subjects' T1 acquisitions to obtain cortical thickness values and characterize the atrophy in every group. All structural T1weighted images were 3D scans, registered in the sagittal plane with an inversion recovery pulse sequence and the following parameters: TR = 2300 ms, TE = 3.0 ms, flip angle = 9.0 degree, voxel size = 1 mm 3 , number of slices = 256. The preprocessing and surface-based morphometry analysis were implemented in CAT12 (http://www.neuro.uni-jena.de/cat), running in SPM12 on MATLAB R2018b.
First, the images were segmented and normalized based on a surface and thickness estimation. Then, cortical thickness data was resampled and smoothed using a 12 mm kernel and merging both hemispheres. Second, we checked data quality, testing the sample homogeneity, and checking orthogonality. Third, the variants were compared with their respective HC subsamples through factorial t-contrasts. The multiple comparisons problem correction was performed with the threshold-free cluster enhancement (TFCE) method (Smith & Nichols, 2009) and the TFCE toolbox (http://www.neuro.uni-jena.de/tfce), which is an extension of SPM12. The cortical thickness comparisons were corrected by the TFCE method with 5000 permutations and a statistical significance level of p < 0.05 (FWE-corrected).
We find the expected atrophy patterns for each FTD variant (Supplementary Figure   1, details in Supplementary Table 1). In bvFTD, cortical thickness decreased strongly in the frontal lobe and the anterior temporal region, as well as in parietal areas but with less affectation (Boxer et al., 2006;Seeley et al., 2009;Whitwell et al., 2009). The CBS showed damage mainly in the areas near the central sulcus, extending to the parietal, temporal, and frontal areas (Seeley et al., 2009;Whitwell et al., 2010). The nfvPPA and PSP groups presented atrophy in the frontal lobe, mainly in superior regions, including premotor and motor areas (Boxer et al., 2006;Lu et al., 2013;Seeley et al., 2009). This damage was lateralized to the right hemisphere in the case of PSP. In patients with svPPA, an anteroposterior gradient of temporal atrophy was detected, extensible to frontal areas like the insular and orbitofrontal cortex, especially in the left hemisphere (Lu et al., 2013;Seeley et al., 2009 Additionally, Supplementary Table 3 shows the statistical power for detecting medium effect sizes on all the group comparisons, with a minimum power of 0.80 as the recommended threshold (Cohen, 1988).

Supplementary data Patient sample harmonization
To reduce possible biases in our samples, the patient groups were matched with two subsamples of HCs. Supplementary Table 4 summarizes the statistical analyses of these comparisons, with two HC subgroups of n = 50. While Supplementary For the discrete variables we used group median comparisons based on a 5000 permutations test, to deal with tied values (Wilcox, 2017). The results are shown as the 0.95 confidence interval for the difference between medians and their respective p-value. For the dichotomic variables, the equality of proportion between groups was analyzed with chi-square and its respective p-value. bvFTD: behavioral variant, CBS: corticobasal syndrome, HC: healthy controls, nfvPPA: nonfluent variant primary progressive aphasia, PSP: progressive supranuclear palsy, svPPA: semantic variant primary progressive aphasia.

Supplementary data Models' performance
Performance details for all models are presented in Supplementary Table 6 and Supplementary Table 7 shows the AUC variability by class through the cross validation. Additionally, we computed all the classifiers with data from scanner 1 (n = 285), see Supplementary Table 8, to avoid any potential biases due to different acquisition parameters. The micro-averages AUC were not statistically different for any classifier using two or one scanner (Supplementary Table 9). Nonparametric tests were implemented to assess statistically significant differences between the ROC curves (Venkatraman, 2000). The variability of the feature importance list was evaluated across nested k-folds to confirm whether the confidence interval of each feature was ranked in the same way as the feature mean. This approach was Result confirmed the stability of the selected features.

Supplementary data Raw connectivity results details
The statistical results of the raw connectivity analysis are shown in Supplementary  Table 10 and Supplementary Figure 5 for functional and structural analysis, respectively.