Trait anxiety is an established risk factor for mental disorders and resting-state functional magnetic resonance imaging (rs-fMRI) could provide an objective marker. Most existing studies focus on functional connectivity, so it is not yet known whether local image-based features hold similar information. We test this on 212 participants from the LEMON dataset. First we demonstrate that band-pass filtering and per-voxel variance normalisation remove the amplitude information, so ALFF is no longer discriminative and fALFF is almost constant. ReHo is based on ranks, so the normalisation does not change it and it remains the only usable image feature. Using ReHo we test linear models, tree models, a support-vector machine and four deep models on the whole 3D map. The problem is framed as regression and as classification at the clinical cut-off. We run the same procedure on two control variables, age and sex. Both controls work. For age we get R² = 0.546 and for sex AUC = 0.786, both with p < 0.001. The sex result also remains inside one age group, where age is almost constant. The features therefore hold at least two kinds of biological information. For trait anxiety the same procedure gives no effect (R² ≈ 0, p = 0.129; AUC = 0.564, p = 0.144). Our correlations are similar to those reported by much larger studies of brain-behaviour associations. This points to a small or absent effect, not a broken procedure. We apply the same pipeline to impulsivity, which carried most of the significant results in the comparison study. That association disappears once we account for age and sex.
|
