The predicted age difference (PAD)—the difference between neuroimaging-predicted brain age and chronological age—has been proposed as a biomarker of individual brain health. Positive PAD indicates accentuated brain aging; negative PAD indicates delayed aging. Prior work has reported PAD deviations in many disorders, but cross-disorder comparisons within a unified framework and the identification of spatial brain patterns and their transcriptomic correlates remain limited. This study aims to systematically compare PAD across nine commonly occurring brain disorders and to examine the spatial brain patterns and biological processes that underlie observed PAD differences.
Structural MRI (sMRI) data were pooled from multiple sources. The training sample comprised healthy controls (HCs) from large consortia (Human Connectome Project [HCP], Genomics Superstruct Project [GSP], and UK Biobank [UKB]) totaling 45,900 HCs. The testing sample included 2,698 patients across nine diagnostic groups plus their matched HCs: ADHD (n = 344), ASD (n = 484), schizophrenia (SZ, n = 152), bipolar disorder (BP, n = 143), major depressive disorder (MDD, n = 258), alcohol use disorder (AUD, n = 155), tobacco use disorder (TUD, n = 144), combined AUD & TUD (A&TUD, sample size reported in source), Alzheimer’s disease (AD, n = 361), and mild cognitive impairment (MCI, n = 657).
Ethical approvals for constituent datasets were reported by the original consortia and local institutions. Some test datasets (psychiatric disorders and addiction) are protected due to privacy and IRB restrictions; other datasets (HCP, GSP, UKB, ADHD-200, ABIDE, ADNI) are publicly available under their respective access policies. The source provides GitHub and Zenodo repositories for code and de-identified minimal data.
Average gray matter volume (GMV) from a high-resolution ROI atlas (augmented Schaefer-1016) was extracted. An XGBoost machine-learning model was trained on age-matched healthy participants and validated by 10-fold cross-validation. The trained model generated individual brain age predictions for testing participants. The PAD was age-corrected and analyzed in linear models that included age, age2, sex, and site as covariates. The PAD difference between each patient group and matched HCs was summarized as Cohen’s d effect sizes with 95% confidence intervals (CIs). Spatial contributions to model predictions were interpreted using Shapley Additive Explanations (SHAP) values; group-by-SHAP interaction terms identified brain regions whose prediction relevance differed between patients and controls.
Across diagnostic groups, PAD was generally higher in patients than in HCs, but effect sizes varied by disorder. The largest effect sizes were observed in dementia: AD showed the highest PAD difference (Cohen’s d = 0.97, 95% CI [0.82, 1.13], p < 0.001), and MCI showed a moderate effect (d = 0.45, 95% CI [0.34, 0.56], p < 0.001). Addiction groups exhibited substantial PAD increases: combined AUD & TUD (A&TUD) d = 0.84 (95% CI [0.44, 1.23], p < 0.001), TUD d = 0.72 (95% CI [0.49, 0.96], p < 0.001), and AUD d = 0.62 (95% CI [0.39, 0.84], p < 0.001). Psychiatric disorders had moderate PAD increases: SZ d = 0.53 (95% CI [0.30, 0.76], p < 0.001), BP d = 0.46 (95% CI [0.22, 0.69], p < 0.001), MDD d = 0.28 (95% CI [0.11, 0.46], p < 0.001). Developmental disorders did not show significant PAD differences: ASD d = 0.06 (95% CI [−0.04, 0.16], p = 0.36) and ADHD d = 0.01 (95% CI [−0.14, 0.15], p = 0.98).
The study also evaluated PAD across age and sex subgroups and tested robustness across datasets, atlas resolutions, and prediction models; the source reports validation steps though full subgroup results are provided in the original article.
Using SHAP-based interaction analyses, the study identified disorder-specific spatial patterns that contributed to increased PAD. A common node across disorders was the prefrontal cortex. For psychiatric disorders, implicated regions centered on frontotemporal networks. Addiction-related PAD mapped to a network configuration including the default-mode network, salience network, putamen, and thalamus. Dementia-related PAD associated with fronto-occipital networks. These spatial patterns indicate that different circuits underlie accentuated brain aging in different disorders, despite shared prefrontal involvement.
The authors linked the spatial PAD difference maps to regional gene expression profiles using publicly available transcriptome resources (Allen Human Brain Atlas and related tools). Genes relatively over- or underexpressed in regions associated with PAD differences were subjected to enrichment analyses. The study reports that disorder-specific PAD-associated genes were enriched in distinct biological processes, suggesting divergent molecular pathways underlie the observed structural aging patterns. Exact gene lists and specific enriched pathways are detailed in the source material.
The study explicitly notes that psychiatric disorders and addiction have high comorbidity and that these potential confounders were not modeled; this may have influenced observed PAD differences. Other limitations discussed in the source include reliance on cross-sectional data for many cohorts and heterogeneity of protected vs publicly available datasets. The source also highlights that transcriptomic associations are correlational and require further validation.
In this large multi-disorder case-control study, accentuated brain aging as measured by PAD was common across several brain disorders but varied in magnitude and spatial anatomy. Dementia showed the largest PAD increases, followed by addiction and psychiatric disorders; developmental disorders showed no consistent PAD differences in the analyzed samples. Disorder-specific brain circuit patterns, with convergent prefrontal involvement, and distinct gene enrichment profiles were identified. The authors suggest these PAD-related brain patterns could serve as neuroimaging biomarkers to probe neural aging mechanisms and ultimately inform clinical decision-making, but further work is needed to address comorbidity confounding and to validate molecular links.