Introduction
Mental and neuropsychiatric disorders are leading contributors to global disability (Arias, Saxena, & Verguet, Reference Arias, Saxena and Verguet2022; Steinmetz et al., Reference Steinmetz, Seeher, Schiess, Nichols, Cao, Servili, Cavallera, Cousin, Hagins, Moberg, Mehlman, Abate, Abbas, Abbasi, Abbasian, Abbastabar, Abdelmasseh, Abdollahi, Abdollahi and Dua2024; Whiteford et al., Reference Whiteford, Degenhardt, Rehm, Baxter, Ferrari, Erskine, Charlson, Norman, Flaxman, Johns, Burstein, Murray and Vos2013), yet objective biomarkers capable of capturing underlying neuropathological processes remain limited (García-Gutiérrez et al., Reference García-Gutiérrez, Navarrete, Sala, Gasparyan, Austrich-Olivares and Manzanares2020). A growing body of evidence suggests that many psychiatric and neurodegenerative conditions involve accelerated biological aging, characterized by structural brain alterations that diverge from normative lifespan trajectories (Hedman et al., Reference Hedman, Van Haren, Schnack, Kahn and Hulshoff Pol2012; Kaufmann et al., Reference Kaufmann, van der Meer, Doan, Schwarz, Lund, Agartz, Alnæs, Barch, Baur-Streubel, Bettella, Beyer, Bøen, Borgwardt, Brandt, Buitelaar, Celius, Cervenka, Conzelmann, Córdova-Palomera and Bertolino2019; Mishra et al., Reference Mishra, Beheshti and Khanna2023). Quantifying these deviations from typical brain aging could provide valuable insights into disease mechanisms, facilitate early diagnosis, and enable monitoring of treatment response. Brain-age modeling, which estimates an individual’s predicted brain age from neuroimaging data, offers a quantitative framework for detecting such deviations. The resulting brain-predicted age difference (brain-PAD) is increasingly recognized as a potential biomarker of disease burden, cognitive decline, functional impairment, and mortality risk across multiple conditions (Cole et al., Reference Cole, Ritchie, Bastin, Valdés Hernández, Muñoz Maniega, Royle, Corley, Pattie, Harris, Wray, Redmond, Marioni, Starr, Cox, Wardlaw, Sharp, Deary and Zhang2018; Cole, Marioni, Harris, & Deary, Reference Cole, Marioni, Harris and Deary2019; Franke & Gaser, Reference Franke and Gaser2019; Kaufmann et al., Reference Kaufmann, van der Meer, Doan, Schwarz, Lund, Agartz, Alnæs, Barch, Baur-Streubel, Bettella, Beyer, Bøen, Borgwardt, Brandt, Buitelaar, Celius, Cervenka, Conzelmann, Córdova-Palomera and Bertolino2019).
Findings across disorders, however, remain inconsistent. Studies in schizophrenia, bipolar disorder, major depressive disorder, and substance use disorders frequently report elevated brain-PAD (Boulos et al., Reference Boulos, Dalwani, Tanabe, Mikulich-Gilbertson, Banich, Crowley and Sakai2016; Dunlop et al., Reference Dunlop, Victoria, Downar, Gunning and Liston2021; Hajek et al., Reference Hajek, Franke, Kolenic, Capkova, Matejka, Propper, Uher, Stopkova, Novak, Kopecek, Spaniel, Alda and Paus2019; Kaufmann et al., Reference Kaufmann, van der Meer, Doan, Schwarz, Lund, Agartz, Alnæs, Barch, Baur-Streubel, Bettella, Beyer, Bøen, Borgwardt, Brandt, Buitelaar, Celius, Cervenka, Conzelmann, Córdova-Palomera and Bertolino2019; Scheffler et al., Reference Scheffler, Ipser, Pancholi, Murphy, Cao, Ottino-González, Thompson, Shoptaw, Conrod, Mackey, Garavan and Stein2024; Schinz et al., Reference Schinz, Schmitz-Koep, Tahedl, Teckenberg, Schultz, Schulz, Zimmer, Sorg, Gaser and Hedderich2023; Shahab et al., Reference Shahab, Mulsant, Levesque, Calarco, Nazeri, Wheeler, Foussias, Rajji and Voineskos2019), whereas evidence in attention-deficit/hyperactivity disorder (ADHD) is mixed (Kurth et al., Reference Kurth, Levitt, Gaser, Alger, Loo, Narr, O’Neill and Luders2022), and neurodegenerative disorders such as Alzheimer’s disease (AD) and frontotemporal dementia (FTD) show large but variable effects (Franke & Gaser, Reference Franke and Gaser2019; Kwon, Yu, & Koh, Reference Kwon, Yu and Koh2024; Lee et al., Reference Lee, Burkett, Min, Senjem, Lundt, Botha, Graff-Radford, Barnard, Gunter, Schwarz, Kantarci, Knopman, Boeve, Lowe, Petersen, Jack and Jones2022; Seo et al., Reference Seo, Thibodeau, Perry, Hua, Sidhu, Sible, Vargas, Gaus, Rabinovici, Rankin, Boxer, Kramer, Rosen, Gorno-Tempini, Grinberg, Huang, DeArmond, Trojanowski, Miller and Seeley2018). These discrepancies may reflect differences in sample size, age range, methodological choices, site variability, or inadequate correction of age-related prediction bias, a well-known limitation in brain-age modeling that can obscure true clinical effects (Cole et al., Reference Cole, Ritchie, Bastin, Valdés Hernández, Muñoz Maniega, Royle, Corley, Pattie, Harris, Wray, Redmond, Marioni, Starr, Cox, Wardlaw, Sharp, Deary and Zhang2018, Reference Cole, Marioni, Harris and Deary2019; Franke & Gaser, Reference Franke and Gaser2019). Few studies have integrated systematic bias correction (Beheshti, Nugent, Potvin, & Duchesne, Reference Beheshti, Nugent, Potvin and Duchesne2019; de Lange & Cole, Reference de Lange and Cole2020; Gupta et al., Reference Gupta, de la Cruz, Rieger, di Giuliano, Gaser, Cole, Breithaupt, Holsen, Eddy, Cetin-Karayumak, Kubicki, Lawson, Miller, Misra, Schumann, Bär and Thomas2025; Liang, Zhang, & Niu, Reference Liang, Zhang and Niu2019), comprehensive interpretability analyses, and cross-disorder comparisons within a single analytical framework (Cole, Reference Cole2020). As a result, it remains unclear whether accelerated brain aging reflects shared transdiagnostic processes or disorder-specific mechanisms. Parallel advances in normative modeling frameworks have begun to address some of these challenges.
Recent large-scale lifespan modeling initiatives, including normative brain charts and centile-based frameworks such as CentileBrain (Bethlehem et al., Reference Bethlehem, Seidlitz, White, Vogel, Anderson, Adamson, Adler, Alexopoulos, Anagnostou, Astle, Auyeung, Ayub, Bae, Ball, Baron-Cohen, Beare, Bedford, Benegal, Beyer and Areces-Gonzalez2022; Ge et al., Reference Ge, Yu, Qi, Fan, Chen, Gao, Haas, New, Boomsma, Brodaty, Brouwer, Buckner, Caseras, Crivello, Crone, Erk, Fisher, Franke, Glahn and Yu2024; Rutherford et al., Reference Rutherford, Kia, Wolfers, Fraza, Zabihi, Dinga, Berthet, Worker, Verdi, Ruhe, Beckmann and Marquand2022), have substantially advanced the characterization of structural brain variability across development and aging. These frameworks provide important population-level reference trajectories and enable benchmarking of normative deviations. However, most prior approaches primarily focus on modeling normative distributions and benchmarking algorithmic performance, rather than integrating bias-corrected brain-age estimation with systematic cross-disorder comparisons and interpretable machine-learning approaches that link regional morphometric features to predicted aging trajectories. In contrast to centile-based normative modeling approaches, the brain-PAD framework provides an interpretable, individual-level deviation metric that is directly comparable across disorders and longitudinal assessments.
A second critical gap arises from the scarcity of lifespan-representative normative models. Most existing studies rely on restricted age ranges or single-site datasets, limiting their ability to detect deviations from true normative aging or to generalize to diverse clinical populations (Bell et al., Bell et al., Reference Bell, Godfrey, Ware, Yeates and Harris2022; Fortin et al., Reference Fortin, Cullen, Sheline, Taylor, Aselcioglu, Cook, Adams, Cooper, Fava, McInnis, Phillips, Trivedi, Weissman, Shinohara and McGrath2018; Mishra et al., Reference Mishra, Beheshti and Khanna2023; Mitchell et al., Reference Mitchell, McDonald, Sharp, Gan, Ponsford, Marquand, Wellington, Law, Shultz and Spitz2025; Rutherford et al., Reference Rutherford, Kia, Wolfers, Fraza, Zabihi, Dinga, Berthet, Worker, Verdi, Ruhe, Beckmann and Marquand2022). Large-scale, multisite datasets encompassing the full human lifespan are necessary to estimate normative neuroanatomical trajectories and quantify pathological deviations robustly (Bethlehem et al., Reference Bethlehem, Seidlitz, White, Vogel, Anderson, Adamson, Adler, Alexopoulos, Anagnostou, Astle, Auyeung, Ayub, Bae, Ball, Baron-Cohen, Beare, Bedford, Benegal, Beyer and Areces-Gonzalez2022; Constantinides et al., Reference Constantinides, Han, Alloza, Antonucci, Arango, Ayesa-Arriola, Banaj, Bertolino, Borgwardt, Bruggemann, Bustillo, Bykhovski, Calhoun, Carr, Catts, Chung, Crespo-Facorro, Díaz-Caneja, Donohoe and Walton2023; Ge et al., Reference Ge, Yu, Qi, Fan, Chen, Gao, Haas, New, Boomsma, Brodaty, Brouwer, Buckner, Caseras, Crivello, Crone, Erk, Fisher, Franke, Glahn and Yu2024; Han et al., Reference Han, Dinga, Hahn, Ching, Eyler, Aftanas, Aghajani, Aleman, Baune, Brak, Filho, Carballedo, Connolly, Couvy-Duchesne, Cullen, Dannlowski, Davey, Dima, Duran and Berger2021; Marek et al., Marek et al., Reference Marek, Tervo-Clemmens, Calabro, Montez, Kay, Hatoum, Donohue, Foran, Miller, Hendrickson, Malone, Kandala, Feczko, Miranda-Dominguez, Graham, Earl, Perrone, Cordova, Doyle and Dosenbach2022; Scheffler et al., Reference Scheffler, Ipser, Pancholi, Murphy, Cao, Ottino-González, Thompson, Shoptaw, Conrod, Mackey, Garavan and Stein2024).
To address these gaps, we constructed a large lifespan neuroimaging dataset comprising 25,425 healthy individuals aged 2–95 years and applied as novel hybrid brain-age prediction model that combines a TabTransformer-based (Huang, Khetan, Cvitkovic, & Karnin, Reference Huang, Khetan, Cvitkovic and Karnin2020) deep-learning architecture with Gaussian process regression (GPR). Using this model, we derived bias-corrected brain-PAD estimates and applied them to independent test sets, including 1,737 patients with seven psychiatric and neurodegenerative disorders and 1,793 matched healthy controls. Our approach incorporated rigorous age-bias correction, harmonization of 288 FreeSurfer-derived morphometric features across 36 imaging sites using ComBat (Fortin et al., Reference Fortin, Cullen, Sheline, Taylor, Aselcioglu, Cook, Adams, Cooper, Fava, McInnis, Phillips, Trivedi, Weissman, Shinohara and McGrath2018; Johnson, Li, & Rabinovic, Reference Johnson, Li and Rabinovic2007), and model interpretability via Shapley Additive Explanations (SHAP) (Lundberg et al., Reference Lundberg, Erion, Chen, DeGrave, Prutkin, Nair, Katz, Himmelfarb, Bansal and Lee2020; Lundberg & Lee, Reference Lundberg and Lee2017) to identify neuroanatomical features that most strongly contribute to brain-age predictions.
The study aimed to (i) establish normative lifespan brain-PAD distributions in healthy individuals; (ii) quantify disorder-specific deviations in brain-PAD across schizophrenia (SZ), major depressive disorder (MDD), bipolar disorder (BD), ADHD, chronic substance use disorder (CSUD), AD, and FTD; (iii) compare regional morphometric differences across disorders to identify transdiagnostic and disorder-specific neuroanatomical patterns via pairwise regional analyses; and (iv) examine the biological plausibility of brain-age estimates using SHAP-derived feature contributions along normative aging trajectories.
We hypothesized that (i) accelerated brain aging would be most pronounced in SZ, MDD, BD, CSUD, AD, and FTD, but not in ADHD; (ii) pairwise comparisons would reveal consistent transdiagnostic involvement of key cortical hubs; and (iii) SHAP analyses would identify biologically plausible, age-sensitive regions, such as ventricular enlargement and cortical thinning, as primary contributors to brain-age predictions.
Methods and materials
Participants
We assembled a large lifespan neuroimaging sample comprising 25,425 healthy individuals aged 2–95 years from 36 sites worldwide, including 24,638 publicly available scans and 787 healthy participants recruited at Jena University Hospital (JUH). This cohort was used exclusively for the development of normative brain-age prediction models. The inclusion of the JUH dataset served two purposes: (i) to improve representation of under-sampled adult age ranges within the lifespan distribution and (ii) to ensure consistent preprocessing quality within a controlled local imaging environment. Inclusion of JUH participants also enabled evaluation of model generalizability across both publicly available and locally acquired MRI data. Detailed demographic and site-level information, including site location, sample size, age distribution, sex composition, and scanner characteristics, is provided in Supplementary Material File 1.
Healthy control (HC) individuals were defined according to the inclusion criteria of the original contributing studies as participants without documented neurological or psychiatric disorders. Public datasets applied dataset-specific screening procedures, typically including structured or semi-structured clinical interviews, medical history review, and exclusion of major neurological or psychiatric diagnoses. JUH participants underwent structured clinical assessment using the Structured Clinical Interview for DSM-IV Axis I Disorders (SCID-I) to confirm the absence of current or past psychiatric or neurological conditions (Supplementary Material File 2, Section 1). Where available, demographic information, including sex and ethnicity, was retained from the original datasets. However, ethnicity information was not consistently reported across all publicly available cohorts, reflecting a common limitation of aggregated multisite neuroimaging studies. To evaluate potential ethnicity- and race-related prediction bias, we performed an additional analysis in an independent subset of healthy participants with available ethnicity and race information (Supplementary Material File 2, Section 6; Figures S6–S7).
To evaluate generalizability of the trained models, two independent HC datasets from the Dallas Lifespan Brain Study (DLBS) (Park et al., Reference Park, Hennessee, Smith, Chan, Chen, Dakanali, Farrell, Liu, Lu, Rofsky, Sun, Tamminga, Moore, Kennedy, Rodrigue and Wig2025) were used for external testing: a cross-sectional cohort (DLBS-CS; n = 270, ages 21–89 years) and a longitudinal cohort including repeated scans obtained at first assessment (FA) and last assessment (LA) ~9 years apart (n = 188). For the longitudinal DLBS cohort, only one scan per time point was used to evaluate prediction stability across time. To examine disorder-specific deviations from normative aging trajectories, we aggregated a clinical sample comprising 1,737 patients across seven diagnostic groups (SZ, MDD, BD, ADHD, AD, FTD, and CSUD) and 1,793 age-matched HCs. Figure 1 summarizes the dataset composition, and Table 1 presents demographic characteristics for each group.
Brain age prediction pipeline using structural MRI neuroimaging and four machine learning models. (a) A pair of circular diagrams summarizes the training and testing datasets. The large outer ring shows the 25,425 healthy control (HC) training scans partitioned by site with counts for total sample, male/female split, and mean age. The smaller ring lists 3,530 testing scans consisting of 1,737 disorder cases and 1,793 matched HCs for 7 disorders. Above it, a concentric ring diagram summarizes the Dallas Lifespan Brain Study (DLBS) test cohort: 270 cross-sectional HCs and 188 longitudinal HCs with first assessment (FA) and last assessment (LA) follow-ups. (b) A schematic workflow shows that raw T1-weighted MRI data are processed through cortical surface extraction and parcellation to derive morphological features. (c) This panel outlines the machine-learning pipeline. The HC dataset is split into training (95%) and validation (5%) sets for fivefold cross-validation. The best-trained model (selected via metrics and bias screening) is then applied to testing data to assess cross-sectional, longitudinal (DLBS-FA and DLBS-LA), and disorder effects. Brain-PAD is depicted, indicating that positive values reflect an ‘older’ brain while negative values reflect a ‘younger’ brain.
Figure 1. Long description
Panel A contains three radial diagrams. The top-left diagram for Training Data H C has a central brain icon labeled 25,425 data. Moving outward, concentric rings show 36 sites with numerical counts for Total Sample, Female, and Male participants. The top-center diagram for D L B S Testing Data H C shows a central brain icon labeled 270 / 188 data, with rings split into Cross-sectional (270) and Longitudinal (188) cohorts, further divided into 120 and 68 sub-groups. The bottom-left diagram for Testing Data Disorder/H C shows a central brain icon labeled 1,737 Disorder and 1,793 H C, with outer segments for B D, F T D, A D, A D H D, C S U D, S Z, and M D D.
Panel B shows a workflow starting with a Raw T 1 w M R I scan, leading to Surface Extraction where morphological cortical features are highlighted. This leads down to three colorful Parcellation Maps showing different brain orientations.
Panel C outlines the Model Training and Validation at the top, where the H C Dataset is split into a 95 percent Training set and a 5 percent Validation set. These feed into M L Algorithms and 5-Fold C V, resulting in a best-trained model. The bottom section, Model Testing, shows this model applied to D L B S-C S, D L B S-F A, D L B S-L A, and various disorder groups. A final line graph plots Chronological Age on the x-axis against Predicted Age on the y-axis. A dashed line represents Brain-P A D equals 0. Data points above the line are labeled Older brain (positive Brain-P A D), and points below are labeled Younger brain (negative Brain-P A D).
Demographic characteristics of healthy controls and patient cohorts
Table 1. Long description
The table is divided into two main sections.
Section 1: Healthy controls (training/validation).
- Training plus validation (36 sites): N = 25,425, Age range 2 to 95, Mean age 45.3 plus or minus 24.8.
- Public sites (35): N = 24,638, Age range 2 to 95, Mean age 45.2 plus or minus 24.7.
- Jena University Hospital (J U H): N = 787, Age range 17 to 85, Mean age 31.9 plus or minus 14.7.
- D L B S-C S: N = 270, Age range 21 to 89, Mean age 60.2 plus or minus 19.0.
- D L B S-F A/L A: N = 188, Age range 21 to 88 or 31 to 97, Mean age 53.9 plus or minus 17.5 or 63.4 plus or minus 16.8.
Section 2: Patient cohorts (test sets) with age-matched healthy controls (H C). Data is listed as Patient (N, Age range, Mean age) followed by Age-matched H C (N, Age range, Mean age).
- S Z: Patient (370, 16 to 67, 37.5 plus or minus 11.9); H C (409, 17 to 68, 36.2 plus or minus 11.3).
- M D D: Patient (472, 18 to 75, 39.1 plus or minus 12.6); H C (373, 18 to 80, 39.4 plus or minus 13.1).
- A D H D: Patient (207, 7 to 50, 14.2 plus or minus 10.3); H C (378, 7 to 50, 15.2 plus or minus 10.0).
- B D: Patient (114, 19 to 86, 42.2 plus or minus 17.0); H C (185, 21 to 89, 37.8 plus or minus 15.8).
- A D: Patient (359, 51 to 98, 74.6 plus or minus 8.0); H C (194, 39 to 92, 67.5 plus or minus 9.7).
- F T D: Patient (121, 42 to 87, 64.9 plus or minus 8.6); H C (172, 39 to 92, 67.0 plus or minus 9.6).
- C S U D: Patient (94, 17 to 50, 28.7 plus or minus 7.9); H C (82, 17 to 48, 29.7 plus or minus 9.7).
Mean ± standard deviation (SD) values for each variable. AD, Alzheimer’s disease; ADHD, attention-deficit/hyperactivity disorder; BD, bipolar disorder; CSUD, chronic substance use disorder; DLBS-CS, Dallas Lifespan Brain Study with cross-sectional scans; DLBS-FA, Dallas Lifespan Brain Study with longitudinal scans at first assessment; DLBS-LA, Dallas Lifespan Brain Study with longitudinal scans at last assessment; FTD, frontotemporal dementia; HC, healthy control subjects; MDD, major depressive disorder; SZ, schizophrenia.
All publicly available datasets were previously de-identified and approved by their respective local ethics committees. JUH participants provided written informed consent, with parental consent obtained for minors. All procedures were conducted in accordance with the Declaration of Helsinki (2013). A comprehensive overview of all contributing datasets, including acquisition details and inclusion criteria, is provided in Supplementary Material File 1.
MRI acquisition and preprocessing
Structural T1-weighted MRI data were processed using the FreeSurfer (version 7.3.2) (Fischl, Reference Fischl2012) cross-sectional ‘recon-all’ pipeline, including standard procedures for motion correction, skull stripping, intensity normalization, topology correction, and cortical surface reconstruction, followed by cortical parcellation (Desikan–Killiany atlas) and subcortical segmentation.
For each participant, we extracted 288 morphometric features, including cortical thickness, cortical volume, surface area, and subjacent white-matter volume (derived from FreeSurfer wmparc labels) across 68 cortical regions defined by the Desikan–Killiany atlas (yielding 4 × 68 = 272 features). We additionally extracted volumetric estimates for 16 subcortical structures: left/right caudate, putamen, pallidum, thalamus, hippocampus, amygdala, nucleus accumbens, and lateral ventricles. The brainstem was excluded because truncation artefacts in several datasets prevented reliable segmentation.
All derived features were visually inspected and subjected to automated quality-control procedures to identify outliers and segmentation errors. Scans failing either procedure were excluded from further analysis.
Because the normative dataset was aggregated from 36 imaging sites using different MRI scanners and acquisition protocols, we applied ComBat harmonization to reduce site-related variability while preserving biologically meaningful variation associated with age and sex. This procedure improves the stability and generalizability of lifespan models by mitigating scanner-specific biases while retaining biologically relevant variability. This harmonization approach enables pooling of heterogeneous datasets without inflating site-specific variance.
Most contributing datasets provided one scan per participant. For datasets containing longitudinal data, only one scan per participant was included in the normative training dataset to avoid bias introduced by repeated measurements. Independent longitudinal data from the DLBS cohort were reserved exclusively for external validation of model stability over time. Harmonized ROI features derived from the normative cohort were then used as input to the brain-age prediction model. Detailed mathematical formulation of the ComBat harmonization procedure and covariate specification are provided in Supplementary Material File 2, Section 3.1.
Brain age estimation and bias correction
Brain age was predicted from the 288 morphometric features using a hybrid TabTransformer-GPR model trained on the large HC cohort. This modeling framework, based on contextual feature learning followed by GPR, was selected for its ability to capture nonlinear lifespan patterns and provide uncertainty estimates. The model architecture, training procedure, hyperparameters, and comparison with other models (XGBoost [Chen & Guestrin, Reference Chen and Guestrin2016], DKL-GPR [Wilson, Hu, Salakhutdinov, & Xing, Reference Wilson, Hu, Salakhutdinov and Xing2016], and TabTransformer [Huang et al., Reference Huang, Khetan, Cvitkovic and Karnin2020] alone) are detailed in Supplementary Material File 2, Section 2. Model performance (mean absolute error [MAE], root mean square error [RMSE], coefficient of determination [R 2], Pearson’s correlation r, and concordance correlation coefficient [CCC]) was quantified using uncorrected predictions to avoid artificially inflating accuracy after bias correction.
Because machine-learning brain-age models typically overestimate younger ages and underestimate older ages, we applied a standard age-bias correction (de Lange & Cole, Reference de Lange and Cole2020; Gupta et al., Reference Gupta, de la Cruz, Rieger, di Giuliano, Gaser, Cole, Breithaupt, Holsen, Eddy, Cetin-Karayumak, Kubicki, Lawson, Miller, Misra, Schumann, Bär and Thomas2025; More et al., Reference More, Antonopoulos, Hoffstaedter, Caspers, Eickhoff and Patil2023; Valdes-Hernandez et al., Reference Valdes-Hernandez, Laffitte Nodarse, Peraza, Cole and Cruz-Almeida2023) derived from the validation set:
where Ω denotes chronological age, and α and β are regression parameters estimated from the validation set. This correction was applied only for inferential analyses (brain-PAD estimation and group differences). The primary outcome measure was the brain-predicted age difference (brain-PAD), defined as:
A positive brain-PAD indicates accelerated brain aging (an older brain age relative to chronological age), whereas a negative brain-PAD reflects delayed or younger-appearing brain age. Brain-PAD values were calculated for all participants in the test datasets. Group-level differences in brain-PAD were statistically evaluated using general linear models (GLMs), with chronological age included as a covariate to control for residual age effects.
Statistical analysis
-
• Healthy controls: normative distributions and longitudinal changes
In HCs, we evaluated median brain-PAD, distributional characteristics, and cross-sectional calibration of predictions across age. Additionally, within-subject aging trajectories were examined in the DLBS longitudinal cohort by assessing agreement between predicted and actual aging over time, operationalized as the mean change in brain-PAD relative to elapsed years between scans
-
• Disorder effects on the brain-PAD
For each disorder, we compared brain-PAD with its age-matched HC group using GLMs with brain-PAD as the dependent variable and diagnosis as the main predictor, controlling for chronological age and sex. We reported effect sizes (Cohen’s d), 95% confidence intervals, and false discovery rate (FDR; Benjamini–Hochberg)-adjusted p-values across disorders. These analyses quantified the magnitude of accelerated brain aging associated with each condition (see Supplementary Material File 2 for details).
-
• Regional morphometric differences between disorders
To characterize whether disorders with elevated brain-PAD exhibited shared or disorder-specific neuroanatomical variation, we conducted pairwise ANCOVAs across all 288 cortical and subcortical morphometric features for the 21 possible disorder–disorder comparisons. Each model included the regional metric as the dependent variable, diagnostic group as the independent variable, and age and sex as covariates. For each comparison, we report the number of significant regions (Pfdr < 0.05), the maximum effect size (η2), and patterns of regional overlap across disorder pairs.
Given the large number of tests, these analyses were designed as exploratory whole-brain characterization of spatial neuroanatomical patterns rather than confirmatory tests of individual regions. FDR correction was applied across the full set of comparisons to control the expected proportion of false positives, and effect sizes are reported to facilitate interpretation of the magnitude of regional differences. These analyses complement global brain-PAD findings by assessing whether disorders showing elevated brain-PAD exhibit structural variation in regions known to change across the lifespan (details of the ANCOVA specification and FDR procedure are provided in Supplementary Material File 2, Section 3.2).
-
• Model interpretability: SHAP analysis
To identify neuroanatomical features that most strongly influenced predicted brain age, we computed SHAP values across the validation set. For each subject, SHAP values quantify whether each cortical or subcortical measure shifted the prediction towards an older or younger brain age relative to the cohort baseline. Feature importance was summarized using mean absolute SHAP values across all validation subjects, supplemented by dependence plots showing the directionality of each feature’s contribution. We additionally compared SHAP-derived feature contributions with normative age trajectories modeled by generalized additive models (GAMs), thereby testing the biological plausibility of the brain-age signal. The full SHAP methodology, sampling procedure, and GAM specification are detailed in Supplementary Material File 2 (Sections 3.3 and 3.4).
Model development, validation, and reporting were conducted in accordance with the TRIPOD-AI (Collins et al., Reference Collins, Moons, Dhiman, Riley, Beam, Van Calster, Ghassemi, Liu, Reitsma, van Smeden, Boulesteix, Camaradou, Celi, Denaxas, Denniston, Glocker, Golub, Harvey, Heinze and Logullo2024) guidelines for transparent reporting of prediction model studies using machine-learning methods.
Results
Model performance and calibration in healthy controls
Across internal and external HC datasets, the TabTransformer-GPR brain-age model demonstrated high predictive accuracy and low systematic bias. In the internal validation cohort, predicted ages closely matched chronological age, with an MAE of 5.40 years and a correlation of r = 0.95 (R 2 = 0.91, CCC = 0.90), computed on uncorrected predictions before age-bias correction. Benchmarking against baseline models (XGBoost, DKL-GPR, and TabTransformer) confirmed that TabTransformer-GPR provided the most favorable balance of accuracy and error calibration across metrics.
When applied to independent datasets from the DLBS, performance declined modestly but remained robust, with MAEs of 6.52 years in the cross-sectional cohort (DLBS-CS), 6.65 years at the first longitudinal assessment (DLBS-FA), and 6.98 years at the last longitudinal assessment (DLBS-LA). Correlations in these independent DLBS cohorts ranged between r = 0.85 and r = 0.91, indicating that predictions generalized well across different scanners, sites, and age ranges. Before age-bias correction, model predictions displayed the expected pattern of overestimating younger individuals and underestimating older individuals. After applying bias correction, the median brain-PAD values were tightly centered around zero: −0.21 years in the internal validation cohort, −0.29 years in DLBS-CS, −0.45 years in DLBS-FA, and + 0.16 years in DLBS-LA (Figure 2a). These findings demonstrate that the corrected brain-PAD metric provides an accurate and unbiased index of normative brain aging across the adult lifespan. Detailed model performance comparisons with TabTransformer, XGBoost, DKL-GPR, and proposed hybrid models are provided in Supplementary Material, Section 4 (Figures S2).
Brain-PAD in healthy controls. (a) Half violin-box plots compare brain-PAD distributions across four healthy groups: the validation set (n = 1,272), DLBS-CS (n = 270), DLBS-FA (n = 188), and DLBS-LA (n = 188). Median brain-PAD values are −0.21, −0.29, −0.45, and + 0.16 years, respectively. The distributions are centered near zero, with slightly more negative bias in the DLBS-FA group and a slight positive shift at LA. (b) The right panel shows the distribution of the within-subject change between the last assessment (LA) and the first assessment (FA) in the DLBS longitudinal cohort. The mean change (red dot) is +9.10 years, closely matching the actual time elapsed between scans (blue dashed line, mean = +9.35 years), indicating that the model accurately tracks aging over time in healthy individuals.
Figure 2. Long description
Panel A, titled Comparison of brain-P A D across healthy groups, features a y-axis labeled brain-P A D (years) ranging from minus 40 to 40. Four half violin-box plots with overlaid scatter points are arranged horizontally. From left to right:
* Validation-Set (n equals 1272) with a median of minus 0.21.
* D L B S-C S (n equals 270) with a median of minus 0.29.
* D L B S-F A (n equals 188) with a median of minus 0.45.
* D L B S-L A (n equals 188) with a median of plus 0.16.
All distributions are centered near the zero dashed line.
Panel B, titled Longitudinal change in predicted brain age (F A minus L A), has a y-axis labeled Difference in predicted age (years) from minus 40 to 40. It shows a single vertical half violin-box plot for the D L B S (F A minus L A) cohort. A horizontal blue dashed line intersects the plot at approximately 9.35 years. A red dot representing the mean change is positioned slightly below the blue line at 9.10 years. Individual data points are scattered to the right of the violin plot.
Longitudinal tracking of brain aging in healthy adults
Longitudinal analyses further confirmed the model’s sensitivity to within-person aging. In the DLBS cohort, individuals underwent two MRI assessments spaced ~9 years apart. The mean increase in predicted age between assessments was +9.10 years (red dot, Figure 2b), aligning closely with the actual elapsed time of +9.35 years (blue dashed line). This near one-to-one correspondence indicates that the model captures the expected temporal progression of brain aging at the group level. Inter-individual variability was observed, as expected in longitudinal data, but the overall distribution of longitudinal change strongly supports the model’s utility for monitoring brain aging within individuals over time.
Disorder-specific deviations in brain-PAD
Comparisons between clinical groups and their age-matched HC groups in HC-adjusted brain-PAD (∆brain-PAD, defined as each individual’s brain-PAD minus the median brain-PAD of the corresponding control group) revealed substantial heterogeneity in patterns of accelerated brain aging. The largest deviations were observed in schizophrenia and frontotemporal dementia, with median ∆brain-PAD values of +7.64 and + 7.61 years, respectively (Figure 3a; both Pfdr < 0.001). These effects corresponded to moderate-to-large effect sizes (Cohen’s d = 0.57 for SZ and d = 0.50 for FTD) and reflect pronounced departures from the normative aging trajectory. Alzheimer’s disease also showed significantly elevated ∆brain-PAD, with a median increase of +3.24 years (d = 0.25, Pfdr < 0.001), consistent with the expected neurodegenerative burden.
Disorder effects and regional differences. (a) For each disorder (ADHD, SZ, BD, MDD, AD, FTD, and CSUD), violin plots show the brain-PAD distribution alongside age-matched controls (gray). Effect sizes (Cohen’s d) and Pfdr quantify the group differences. SZ (d = 0.57), AD (d = 0.25), FTD (d = 0.50), and CSUD (d = 0.62) show pronounced positive shifts (older-appearing brains) relative to controls, while ADHD shows no significant difference (d = 0.04, Pfdr = 0.62). (b) A chord diagram displays effect sizes (η2) for pairwise differences among the seven disorders. Thicker chords and warmer colors indicate larger effect sizes; the legend ranges from 0 to 0.15.
Figure 3. Long description
Panel a is a vertical series of violin plots and box plots comparing brain-P A D in years for seven disorders against gray control distributions. The x-axis ranges from minus 50 to 55. A vertical dashed line marks the zero point. From top to bottom:
* A D H D: Red distribution, d equals 0.04, p equals 0.621, n equals 207.
* S Z: Gold distribution, d equals 0.57, p equals 3.7 times 10 super minus 12, n equals 370.
* B D: Pink distribution, d equals 0.31, p equals 0.0286, n equals 114.
* M D D: Orange distribution, d equals 0.16, p equals 0.0141, n equals 472.
* A D: Purple distribution, d equals 0.25, p equals 1.6 times 10 super minus 04, n equals 359.
* F T D: Light blue distribution, d equals 0.5, p equals 1.7 times 10 super minus 05, n equals 121.
* C S U D: Teal distribution, d equals 0.62, p equals 0.00536, n equals 94.
Panel b is a circular chord diagram showing pairwise effect sizes eta-squared. The perimeter is divided into segments for S Z, A D, A D H D, B D, C S U D, F T D, and M D D. Internal ribbons connect these segments; thicker, darker red ribbons indicate higher effect sizes, while thinner yellow ribbons indicate lower values. A color scale at the bottom right indicates effect sizes from 0 to 0.15.
Moderate elevations in ∆brain-PAD were observed in bipolar disorder (+2.08 years, d = 0.31, Pfdr < 0.05) and chronic substance use disorder (+2.91 years, d = 0.62, Pfdr < 0.01). Major depressive disorder exhibited a smaller but still significant shift, with a median increase of +2.76 years (d = 0.16, Pfdr < 0.05). In contrast, attention-deficit/hyperactivity disorder showed no significant difference from age-matched controls; median ∆brain-PAD values differed by only −0.02 years (d = 0.04, Pfdr > 0.05). This pattern aligns with prior findings suggesting that ADHD is characterized more by neurodevelopmental delay than by accelerated aging. Collectively, these results show that accelerated brain aging is a prominent feature of several psychiatric and neurodegenerative disorders but is not uniformly present across diagnostic categories.
Regional neuroanatomical differences between disorders
To contextualize global brain-PAD findings, we examined regional morphometric differences across all 21 possible disorder–disorder comparisons. Across the 288 cortical and subcortical regions tested, 3,313 significant differences were identified after FDR correction. Effect sizes varied across disorder pairs, with contrasts involving SZ, FTD, AD, BD, and MDD generally showing more widespread regional divergence (Figure 3b). Comparisons involving ADHD showed fewer significant regional differences, consistent with the absence of global brain-PAD elevation in this group.
Several cortical and subcortical regions appeared repeatedly across multiple disorder pairs. In particular, left precuneus thickness, left supramarginal gyrus thickness, and left hippocampal volume emerged as significant features across numerous comparisons, including SZ versus BD, AD versus FTD, CSUD versus MDD, and ADHD versus SZ.
At the same time, disorder-specific patterns emerged; for example, AD and FTD showed distinct temporal and limbic profiles, whereas SZ and BD differed more prominently in frontoparietal regions. These findings indicate that structural variation across disorders reflects a combination of shared and disorder-specific neuroanatomical characteristics, consistent with the heterogeneous biological mechanisms underlying neuropsychiatric conditions. Detailed regional results, including chord-diagram visualizations and complete pairwise statistics, are provided in Supplementary Material File 2, Section 5 (Figures S3–S5 and Table S1), and full numerical results for each region–pair combination are provided in Table S5 of Supplementary Material File 1.
SHAP-derived feature importance and biological interpretability
To evaluate whether model predictions reflected biologically meaningful aspects of brain structure, we examined feature contributions using the SHAP algorithm. Across the HC validation cohort (n = 1,272), four neuroanatomical features consistently exerted the strongest influence on predicted brain age: left precentral volume (mean SHAP = 3.20), left isthmus cingulate area, left isthmus cingulate volume, and left lateral ventricle volume (Figure 4a). These features stood out clearly from the remaining 288 cortical and subcortical measures, most of which had near-zero mean SHAP values (< 0.5), indicating minimal contribution to the model’s predictions. SHAP dependence plots provided granular insight into these relationships (Figure 4b): higher ventricular volumes robustly increased predicted age, consistent with known age-related ventricular expansion, whereas lower cortical volumes and surface areas in the precentral and cingulate regions shifted predictions toward older age. Notably, the left isthmus cingulate volume exhibited a more complex nonlinear pattern, with low volumes associated with younger predicted age, intermediate volumes exerting little influence, and higher volumes in older individuals contributing to older predicted age. These patterns align with established trajectories of cortical thinning, volume loss, and ventricular enlargement across the lifespan.
Feature importance and age-related trajectories. (a) A bar chart ranks features by mean absolute SHAP value. The four most influential features – left precentral volume, left isthmus cingulate area, left isthmus cingulate volume, and left lateral ventricle volume – have SHAP values around 2.6–3.2. Smaller contributions come from right pericalcarine thickness, right temporal pole white-matter volume, right entorhinal thickness, and left middle temporal area. (b) Scatter plots show how SHAP values change with feature magnitude. For the left precentral and isthmus cingulate area, larger volumes are associated with negative SHAP values (predicting younger age), whereas larger left isthmus cingulate volumes and lateral ventricle volumes produce positive SHAP values (predicting older age). (c) Four scatter plots illustrate the relationship between chronological age and each top-ranked feature in healthy controls. Left precentral and isthmus cingulate measures decrease slightly with age, while left lateral ventricle volume increases markedly with age.
Figure 4. Long description
Panel a is a bar chart titled Mean absolute S H A P value. The y-axis represents Mean absolute S H A P value from 0.0 to 3.0. Eight bars are ranked from highest to lowest. The top four features are Left Precentral Volume at 3.20, Left Isthmuscingulate Area at 3.02, Left Isthmuscingulate Volume at 2.62, and Left Lateral Ventricle at 2.61. A break in the x-axis separates these from four lower-ranking features: Right Pericalcarine Thickness at 0.23, Right Temporalpole W M at 0.23, Right Entorhinal Thickness at 0.21, and Left Middletemporal Area at 0.19.
Panel b contains four scatter plots showing S H A P value on the y-axis against feature magnitude on the x-axis. A legend indicates that positive values predict older age and negative values predict younger age.
* Top-left: Left Precentral Volume (mm cubed) shows a strong negative linear correlation.
* Top-right: Left Isthmuscingulate Area (mm squared) shows a strong negative linear correlation.
* Bottom-left: Left Isthmuscingulate Volume (mm cubed) shows a strong positive linear correlation.
* Bottom-right: Left Lateral Ventricle Volume (mm cubed) shows a strong positive non-linear correlation.
Panel c contains four scatter plots showing the relationship between Chronological Age (years) on the x-axis and brain measurements on the y-axis.
* Left Precentral Volume: Shows a steady linear decrease from age 0 to 90.
* Left Isthmuscingulate Area: Shows a slight, nearly flat linear decrease.
* Left Isthmuscingulate Volume: Shows a slight downward curve that plateaus after age 50.
* Left Lateral Ventricle: Shows a marked exponential increase in volume starting around age 50.
GAMs further validated these patterns, showing that these high-SHAP regions follow biologically expected developmental and aging trajectories across the full healthy control cohort (n = 25,425, age range 2–95 years) (Figure 4c). Cortical regions demonstrated a nonlinear decline across the lifespan, with rapid reductions in youth, relative stabilization in midlife, and accelerated atrophy in older adulthood, whereas ventricular volumes increased slowly until midlife and more rapidly thereafter. The close correspondence between SHAP-derived importance, dependence patterns, normative aging curves, and the disorder-related deviations described above supports the biological validity of the brain-age signal, indicating that predictions are driven by meaningful neuroanatomical variation rather than dataset artefacts.
Discussion
The present study used a large lifespan neuroimaging cohort to characterize normative brain aging and to quantify deviations across seven psychiatric and neurodegenerative disorders. Several key findings emerged. First, after rigorous harmonization and age-bias correction, brain-PAD values in HCs were tightly centered around zero across all age groups and remained stable across independent test datasets. Second, longitudinal analyses showed that predicted brain age closely tracked within-person aging over nearly a decade, with the mean increase in predicted age approximating the mean elapsed follow-up interval, supporting the temporal validity of the model. Third, six of the seven disorders examined demonstrated significantly elevated brain-PAD, with the largest effects observed in schizophrenia and frontotemporal dementia, whereas ADHD showed no evidence of accelerated aging. Finally, regional and interpretability analyses indicated that both shared and disorder-specific neuroanatomical alterations contribute to elevated brain-PAD, with repeated involvement of the precuneus, supramarginal gyrus, and hippocampus across multiple diagnostic comparisons. These regional patterns should be interpreted as convergent supporting evidence rather than direct mechanistic explanations of accelerated aging. Overall, the results suggest that brain-PAD captures distributed structural differences across disorders that align with known age-related brain changes.
Accelerated brain aging as a transdiagnostic feature
A central finding of this study is that accelerated brain aging appears across a diverse range of neuropsychiatric disorders with varying magnitudes. SZ and FTD showed the most pronounced elevations in brain-PAD, consistent with their well-established structural abnormalities and progressive neuropathological features. In SZ, converging evidence supports a combination of early neurodevelopmental divergence and subsequent neuroprogressive changes involving cortical thinning, ventricular enlargement, and synaptic dysfunction (Birnbaum & Weinberger, Reference Birnbaum and Weinberger2017; Faludi & Mirnics, Reference Faludi and Mirnics2011). The observed elevation of ~7 years in ∆brain-PAD aligns with prior work (Haas et al., Reference Haas, Ge, Sanford, Modabbernia, Reichenberg, Whalley, Kahn and Frangou2022; Kaufmann et al., Reference Kaufmann, van der Meer, Doan, Schwarz, Lund, Agartz, Alnæs, Barch, Baur-Streubel, Bettella, Beyer, Bøen, Borgwardt, Brandt, Buitelaar, Celius, Cervenka, Conzelmann, Córdova-Palomera and Bertolino2019; Koutsouleris et al., Reference Koutsouleris, Davatzikos, Borgwardt, Gaser, Bottlender, Frodl, Falkai, Riecher-Rössler, Möller, Reiser, Pantelis and Meisenzahl2014) showing increased apparent brain age and supports the hypothesis that SZ reflects not only aberrant neurodevelopment but also an acceleration of aging-related processes across adulthood.
Similarly, the large deviations in FTD, as expected for a neurodegenerative disorder, reflect its hallmark features such as cortical atrophy, ventricular enlargement, and degeneration in frontal and temporal brain networks, diverging from the normative trajectory (Feng et al., Reference Feng, Lipton, Yang, Small and Provenzano2020; Lee et al., Reference Lee, Burkett, Min, Senjem, Lundt, Botha, Graff-Radford, Barnard, Gunter, Schwarz, Kantarci, Knopman, Boeve, Lowe, Petersen, Jack and Jones2022; Seo et al., Reference Seo, Thibodeau, Perry, Hua, Sidhu, Sible, Vargas, Gaus, Rabinovici, Rankin, Boxer, Kramer, Rosen, Gorno-Tempini, Grinberg, Huang, DeArmond, Trojanowski, Miller and Seeley2018). The magnitude of brain-PAD elevation observed in this group indicates that brain-age models are sensitive to disease-related structural decline and can differentiate between patterns of accelerated neurodegeneration and normal aging.
AD exhibited moderate but still significant acceleration in brain-PAD. This pattern likely reflects the interaction between early focal degeneration, particularly in medial temporal structures, and the broader involvement of cortical and subcortical regions that emerges later in the disease progression (Franke & Gaser, Reference Franke and Gaser2019; Lee et al., Reference Lee, Burkett, Min, Senjem, Lundt, Botha, Graff-Radford, Barnard, Gunter, Schwarz, Kantarci, Knopman, Boeve, Lowe, Petersen, Jack and Jones2022). The degree of deviation observed here is consistent with previous lifespan studies (Kaufmann et al., Reference Kaufmann, van der Meer, Doan, Schwarz, Lund, Agartz, Alnæs, Barch, Baur-Streubel, Bettella, Beyer, Bøen, Borgwardt, Brandt, Buitelaar, Celius, Cervenka, Conzelmann, Córdova-Palomera and Bertolino2019) and further underscores the clinical value of brain-PAD in capturing the disease burden associated with neurodegenerative conditions.
Moderate elevations in BD, MDD, and CSUD suggest these conditions also involve structural changes consistent with accelerated aging. In BD and chronic depression, chronic inflammation, oxidative stress, mitochondrial dysfunction, and recurrent mood episodes have been proposed as mechanisms contributing to cumulative brain burden (Christman et al., Reference Christman, Bermudez, Hao, Landman, Boyd, Albert, Woodward, Shokouhi, Vega, Taylor and Andrews2020; Kim, Santos, Gage, & Marchetto, Reference Kim, Santos, Gage and Marchetto2017; Scaini et al., Reference Scaini, Rezin, Carvalho, Streck, Berk and Quevedo2016). Likewise, CSUD is associated with neurotoxic effects, including excitotoxicity, disrupted white-matter integrity, and alterations in frontostriatal circuits (Crunelle et al., Reference Crunelle, Kaag, Van Den Munkhof, Reneman, Homberg, Sabbe, van den Brink and Van Wingen2015; Hampton, Hanik, & Olson, Reference Hampton, Hanik and Olson2019). The observed brain-PAD increases in CSUD were comparable in magnitude to those in SZ, reinforcing concerns about long-term neurobiological consequences of substance use and emphasizing the need for early intervention.
Conversely, ADHD did not exhibit elevated brain-PAD, suggesting a distinct pathophysiological profile. While ADHD is characterized by atypical neurodevelopmental trajectories, delayed maturation, and alterations in frontoparietal networks, its structural features do not map onto accelerated aging (Ágrez et al., Reference Ágrez, Vakli, Weiss, Vidnyánszky and Bunford2025; Kaufmann et al., Reference Kaufmann, van der Meer, Doan, Schwarz, Lund, Agartz, Alnæs, Barch, Baur-Streubel, Bettella, Beyer, Bøen, Borgwardt, Brandt, Buitelaar, Celius, Cervenka, Conzelmann, Córdova-Palomera and Bertolino2019). This divergence highlights that brain-PAD differentiates between developmental delay and neuroprogressive processes. Overall, these cross-disorder patterns indicate that accelerated brain aging is a unifying feature of conditions involving chronic illness burden, metabolic dysregulation, and neuroprogressive or neurodegenerative mechanisms, potentially reflecting cumulative effects of disease course, treatment exposure, and environmental stressors.
Regional analyses identify shared vulnerability hubs across disorders
While brain-PAD provides a global metric of aging-related deviation from normative aging trajectories, the regional analyses provide complementary spatial context regarding how structural differences are distributed across disorders. Across the 21 disorder–disorder comparisons, several cortical and subcortical regions emerged repeatedly, particularly the precuneus, supramarginal gyrus, and hippocampus. These regions form key nodes in networks supporting memory, executive function, social cognition, integrative cognitive processing, and self-referential processing. Furthermore, these areas are known to exhibit significant structural variation across the lifespan and are consistently reported in studies of normative aging and neurodegeneration (Bethlehem et al., Reference Bethlehem, Seidlitz, White, Vogel, Anderson, Adamson, Adler, Alexopoulos, Anagnostou, Astle, Auyeung, Ayub, Bae, Ball, Baron-Cohen, Beare, Bedford, Benegal, Beyer and Areces-Gonzalez2022; Fjell & Walhovd, Reference Fjell and Walhovd2010; Kaufmann et al., Reference Kaufmann, van der Meer, Doan, Schwarz, Lund, Agartz, Alnæs, Barch, Baur-Streubel, Bettella, Beyer, Bøen, Borgwardt, Brandt, Buitelaar, Celius, Cervenka, Conzelmann, Córdova-Palomera and Bertolino2019).
Importantly, these findings should be interpreted as descriptive spatial patterns rather than direct evidence that specific morphometric alterations cause accelerated brain aging. Because brain-age models integrate distributed morphometric information across the entire brain, regional differences represent spatial correlates of global deviations rather than isolated causal mechanisms. Instead, they provide complementary information indicating that disorders showing elevated brain-PAD also exhibit structural differences in regions commonly implicated in lifespan neuroanatomical change. The convergence between regional differences, SHAP feature-importance patterns, and known lifespan trajectories suggests that brain-PAD reflects biologically meaningful structural variation while acknowledging that multiple interacting processes likely contribute to observed deviations.
At the same time, distinct regional profiles were evident across disorders. For example, AD showed prominent medial temporal and hippocampal alterations, whereas FTD exhibited stronger frontal and limbic signatures. SZ and BD differed primarily in frontoparietal and cingulate regions, reflecting differences in cognitive control, emotional regulation, and functional network integration. CSUD demonstrated patterns involving cortical thinning and ventricular enlargement consistent with prior neuroimaging findings. These spatial differences highlight that accelerated brain aging is not a uniform process but reflects heterogeneous structural alterations across disorders. Importantly, the presence of regional differences in disorder–disorder comparisons indicates that structural variation extends beyond simple deviation from a healthy baseline and reflects distinct neurobiological profiles across conditions. These findings support the interpretation that brain-PAD captures distributed structural variation across multiple brain systems rather than isolated regional abnormalities.
Biological plausibility of brain-age predictions
A critical strength of this study is the integration of predictive modeling with interpretability analyses, which demonstrated that model predictions rely on neuroanatomical features known to be sensitive to aging. SHAP analyses highlighted the prominent role of cortical thinning and ventricular enlargement, canonical markers of aging and neurodegeneration (Aycheh et al., Reference Aycheh, Seong, Shin, Na, Kang, Seo and Sohn2018; Boulos et al., Reference Boulos, Dalwani, Tanabe, Mikulich-Gilbertson, Banich, Crowley and Sakai2016). For instance, increased lateral ventricular volume strongly predicted older brain age and showed clear trajectories of nonlinear expansion across the lifespan. Similarly, reduced volume and surface area in the precentral gyrus and isthmus cingulate predicted older age and reflected expected patterns of cortical decline. These SHAP-derived contributions mirrored normative trajectories derived from generalized additive models, which showed nonlinear cortical decline and ventricular expansion across ages 2–95 years in the full HC lifespan. This convergence between SHAP-derived model importance, normative aging curves, and disorder-related abnormalities confirms that the algorithm captures biologically meaningful information rather than scanner artifacts or nonbiological statistical noise. This interpretability is crucial for clinical translation, providing transparency regarding which brain regions contribute most to apparent aging and how these regions differ across neuropsychiatric disorders.
Clinical implications
These findings have several important clinical implications. First, brain-PAD may serve as a complementary biomarker in psychiatric research, providing an objective index of deviation from normative brain-aging trajectories. Prior studies have reported associations between elevated brain-PAD and cognitive impairment (Gupta et al., Reference Gupta, de la Cruz, Rieger, di Giuliano, Gaser, Cole, Breithaupt, Holsen, Eddy, Cetin-Karayumak, Kubicki, Lawson, Miller, Misra, Schumann, Bär and Thomas2025; Kaufmann et al., Reference Kaufmann, van der Meer, Doan, Schwarz, Lund, Agartz, Alnæs, Barch, Baur-Streubel, Bettella, Beyer, Bøen, Borgwardt, Brandt, Buitelaar, Celius, Cervenka, Conzelmann, Córdova-Palomera and Bertolino2019), functional decline (Crunelle et al., Reference Crunelle, Kaag, Van Den Munkhof, Reneman, Homberg, Sabbe, van den Brink and Van Wingen2015; Yuan et al., Reference Yuan, Zhu, Ren, Yuan, Gao, Zhang, Li, Meng, Gong, Lui, Qiu and Zhang2018), and increased mortality risk (Cole et al., Reference Cole, Ritchie, Bastin, Valdés Hernández, Muñoz Maniega, Royle, Corley, Pattie, Harris, Wray, Redmond, Marioni, Starr, Cox, Wardlaw, Sharp, Deary and Zhang2018). While the present study did not directly assess clinical outcomes, the observed cross-disorder differences indicate that brain-PAD captures biologically meaningful variation across neuropsychiatric conditions and may provide a useful quantitative phenotype for future longitudinal and clinical investigations. Second, the observed differences in brain-PAD across disorders suggest that accelerated brain aging may contribute to differentiating conditions with superficially overlapping symptoms but distinct neurobiological trajectories. For instance, the absence of elevated brain-PAD in ADHD contrasts with the pronounced deviations observed in schizophrenia and neurodegenerative disorders, consistent with the distinction between neurodevelopmental delay and neuroprogressive processes. Such differences highlight the potential utility of brain-age metrics as complementary tools for characterizing heterogeneity across psychiatric conditions. Third, longitudinal tracking of brain-PAD may help characterize individual variability in aging trajectories over time. Although longitudinal clinical outcomes were not examined in the present study, the stability of brain-PAD across repeated measures in the independent DLBS cohort suggests that the metric may provide a reliable marker for future studies investigating disease progression or treatment effects. Finally, the identification of transdiagnostic regions frequently implicated across disorders suggests that shared neurobiological mechanisms may contribute to structural brain variation across psychiatric and neurodegenerative conditions.
Limitations of the study
Several limitations of our study should be acknowledged. First, despite harmonizing data across 36 sites using ComBat, residual scanner- or population-related variability may persist; and large multisite datasets improve generalizability but also introduce heterogeneity that is difficult to eliminate fully. Second, although demographic variables such as sex were available across most datasets, ethnicity information was not consistently reported across all contributing cohorts, reflecting a common limitation of aggregated legacy neuroimaging datasets. Consequently, ethnicity could not be systematically modeled as a covariate during model development. Nevertheless, an independent evaluation using 1,365 healthy participants with available ethnicity and race information found no evidence of systematic differences in brain-PAD across the available demographic groups (Supplementary Material File 2, Section 6; Figures S6 and S7). These findings should be interpreted cautiously because demographic information remained incomplete across the aggregated datasets, and some racial groups were relatively small. Future studies using larger, demographically balanced cohorts will be important to further evaluate model performance across diverse populations. Third, while FreeSurfer is not optimized for young children, a prior work has demonstrated acceptable segmentation accuracy (Ghosh et al., Reference Ghosh, Kakunoori, Augustinack, Nieto-Castanon, Kovelman, Gaab, Christodoulou, Triantafyllou, Gabrieli and Fischl2010; Zughayyar et al., Reference Zughayyar, Bauer, Güttler, de Almeida Marcelino, Kühne, Buss, Heim, Aigner, Tietze and Dell’Orco2025), and the consistency of the observed age-related trajectories with established lifespan patterns supports the validity of the pediatric data. Fourth, smaller sample sizes in the FTD and CSUD groups limited statistical power to detect subtle within-disorder and regional subgroup differences. Fifth, the lack of uniform clinical data (such as medication, illness duration, and symptom severity) across datasets restricted our ability to directly link brain-PAD with clinical characteristics. Finally, although longitudinal validation in the DLBS cohort was robust, extended follow-up in patient groups and the integration of multimodal data (such as fMRI and PET) are needed to comprehensively characterize accelerated aging trajectories.
Conclusion
This study establishes brain-PAD as a robust, transdiagnostic index of biological aging across the lifespan. Using a hybrid TabTransformer–GPR model trained on a large, harmonized, age-diverse normative sample, we demonstrated that brain-PAD aligns with normative trajectories, captures within-person longitudinal change, and reveals clinically meaningful deviations (MAE ≈ 5–7 years). Accelerated brain aging was most pronounced in schizophrenia, frontotemporal dementia, and Alzheimer’s disease, with more moderate elevations in bipolar disorder, major depressive disorder, and chronic substance use disorder, whereas ADHD showed no deviation from normative aging, underscoring its distinct neurodevelopmental profile. Interpretability analyses indicated that these deviations reflect biologically plausible, disorder-specific structural alterations. These findings highlight brain-PAD as a clinically informative marker reflecting both shared and distinct neuropathological processes.
Supplementary material
The supplementary material for this article can be found at http://doi.org/10.1017/S003329172610587X.
Acknowledgments
The authors would like to thank the 36 data providers for publicly or privately sharing their healthy control, neurodegenerative, and psychiatric disorder datasets, which greatly facilitated this research.
Funding statement
Andy Schumann and Feliberto de la Cruz received funding from the Interdisciplinary Center for Clinical Research, Jena University Hospital, Germany (AMSP17, MSP19), and the German Research Foundation (SCHU3432/2-1, CR 994/2-1). The funding agencies had no role in the study design; the collection, analysis, or interpretation of data; the writing of the report; or the decision to submit the article for publication.
Competing interests
The authors declare no competing interests.
Ethical standards
The authors assert that all procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation and with the Helsinki Declaration of 1975, as revised in 2008. All participants (and legal guardians for participants under the age of 18 years) gave informed written consent. The study was approved by the local ethics committee of Jena University Hospital and conducted in accordance with the Declaration of Helsinki (2013). Informed consent was obtained from adult participants, and for those under 18 years, parental/guardian consent, along with child assent, was secured.
Open access
