How Well Does Structural MRI Classify Autism? A Leakage-Free, Site-Aware and Externally Validated Machine-Learning Evaluation in Two ABIDE Cohorts

Regularised logistic regression on FreeSurfer structural MRI features distinguishes ASD from TD in ABIDE II only modestly, reaching AUC 0.60 [0.56, 0.64] under leakage-free, site-aware validation.

• Why it matters: Prior structural MRI studies may overstate autism classification accuracy due to small samples, single splits, site effects, and information leakage, leaving uncertainty about real-world generalization.
• What they did: Using ABIDE II (804 participants; 341 ASD; 12 sites), the authors trained multiple classifiers on morphometry features with confound regression and site-wise standardisation learned within training folds, then evaluated via nested 10-fold CV, leave-one-site-out, and external testing on ABIDE I.
• The result: External ABIDE I AUCs were 0.52–0.58 and leave-one-site-out pooled AUCs were 0.54–0.61 with large site variation; permutation testing supported only chance-level discrimination.

AI summary of the abstract · may contain errors · not medical advice · read the paper.