Figure 2
Test-retest reliability of FCD measures in pre-HD. (a) Correlation of test and retest FCD within subjects. For each subject’s baseline scan, we computed the correlation (across brain voxels) of the FCD map with that of the follow-up FCD map of all 153 subjects and ranked them. The histogram shows that, for 67% of subjects, the top-ranked correlation with follow-up FCDs was with their own follow-up, identifying them. Red line: 0.65% (i.e., 1/153) chance level. X axis is truncated at 10 for visibility. (b) Multivariate detection of group differences. ROC curve of a full-brain classifier of pre-HD vs controls (LOSO-CV). AUC: 0.64 (p = 0.00082, permutation test). Group differences are detectable but mild, compatible with heterogeneity. (c) Test-retest reliability of group differences. Correct and incorrect classifications are highly reliable, suggesting they are determined by consistent heterogeneity. Left: Matrix indicating probability of classification outcome of a follow-up FCD sample, given that of the baseline sample. Right: Chance level matrix (permutation test, see Methods). Both hits and errors are reliable within subjects. *: p < 10−3, **: p < 10−5 (group-preserving two-sided permutation test, see Methods) TN: true negative, FP: false positive, FN: false negative, TP: true positive. (d) Univariate FCD group differences. Single cluster on the globus pallidus and interface with the putamen (two-sided Mann–Whitney U test, FDR q < 0.05). Red: pre-HD > controls, Blue: pre-HD < controls. Units: logarithm of corrected p-value. (e) Mean maps of classifier weights reveal a bilateral globus pallidus cluster. Weights shown at an arbitrary threshold of 3, in standard units across voxels.
