Reliability of Brain MRI and DTI Patterns for Distinguishing Dp140 Subgroups in Duchenne Muscular Dystrophy Using Cross-Validated Classification

Background. Boys with Duchenne muscular dystrophy (DMD) who lack the brain dystrophin isoform Dp140 show larger cognitive and imaging differences than those who keep it, but published MRI and diffusion (DTI) cohorts hold only 18 to 60 patients, and it is unknown how reliably a subgroup difference can be detected or classified at that size. Aim. To test whether region-level MRI/DTI measures differ between Dp140+ and Dp140− boys and to quantify how far a cross-validated classifier can be trusted at small n. Methods. A fixed analysis order: group statistics with effect sizes first, then a regularised classifier inside one pipeline in which every label-dependent step is refitted within each fold (nested cross-validation), permutation tests, learning curves, power and feature stability. A matched dataset is not yet secured, so the pipeline was exercised on published summary statistics, on simulated data, and on real regional cortical thickness from ABIDE (autism versus controls, n = 218). Results. In the published 12 versus 12 Dp140 cohort no brain volume differed after correction (white matter g = 0.78, Holm p = 0.31). Selecting features before cross-validation inflated AUC on pure noise to 0.72 at n = 24. On real ABIDE data the classifier beat the permutation null overall (AUC 0.62, p ≤ 0.005 with 200 permutations), but at n = 24 to 60 it did so in only 13% to 23% of random subsamples. Conclusion. Current cohort sizes are too small for a reliable Dp140 classifier; for the small real effect in ABIDE (median |d| = 0.20) about 150 children were needed for 75% power in subsamples of one 218-child sample, which is an optimistic figure; the number for Dp140 depends on the true effect, which is unknown. Requests for DMD data are pending.