Multimodal MRI marker of cognition explains the association between cognition and mental health in the UK Biobank

Multimodal MRI marker of cognition explains the association between cognition and mental health in the UK Biobank eLife Assessment This valuable work advances our understanding of the relationship between multimodal magnetic resonance imaging (MRI) measures, cognition, and mental health. Compelling use of statistical learning techniques in UK Biobank data shows that 48% of the variance between an 11-task derived g-factor and imaging data can be explained. Overall, this paper contributes to the study of brain-behaviour relations and will be of interest for both its methods and its findings on how much variance in g can be explained. https://doi.org/10.7554/eLife.108109.3.sa0Valuable: Findings that have theoretical or practical implications for a subfield - Landmark - Fundamental - Important - Valuable - Useful Compelling: Evidence that features methods, data and analyses more rigorous than the current state-of-the-art - Exceptional - Compelling - Convincing - Solid - Incomplete - Inadequate During the peer-review process the editor and reviewers write an eLife Assessment that summarises the significance of the findings reported in the article (on a scale ranging from landmark to useful) and the strength of the evidence (on a scale ranging from exceptional to inadequate). Learn more about eLife Assessments Abstract Cognitive dysfunction often co-occurs with psychopathology. Advances in neuroimaging and machine learning have led to neural indicators that predict individual differences in cognition with reasonable performance. We examined whether these indicators explain the relationship between cognition and mental health in the UK Biobank (n>14,000). Using machine learning, we quantified the covariation between cognition and 133 mental health indices and derived neural indicators of cognition from 72 neuroimaging phenotypes across diffusion-weighted MRI (dwMRI), resting-state functional MRI (rsMRI), and structural MRI (sMRI). With commonality analyses, we investigated how much of the cognition–mental health covariation is captured by each indicator and neural indicators combined within and across MRI modalities. The predictive association between mental health and cognition was at r=0.3. Neuroimaging captured 2.1 to 25.8% of the cognition-mental health covariation. Combining phenotypes within modalities improved the explanation to 25.5% for dwMRI, 29.8% for rsMRI, and 31.6% for sMRI, and combining them across modalities enhanced the explanation to 48%. We present an integrated approach to derive multimodal MRI markers of cognition that can be transdiagnostically linked to psychopathology, demonstrating that the predictive ability of neural indicators extends beyond the prediction of cognition itself, enabling us to capture cognition-mental health covariation. Introduction Cognition and mental health are closely intertwined (Iosifescu, 2012). Cognitive dysfunction is present in various mental illnesses, including anxiety (Gulpers et al., 2022; Nyberg et al., 2021), depression (Kriesche et al., 2023; Richardson and Adams, 2018; Wen et al., 2022), and psychotic disorders (Chavez-Baldini et al., 2023; Fusar-Poli et al., 2012; Guo et al., 2019; Lindgren et al., 2020; Mesholam-Gately et al., 2009; Semkovska et al., 2019). National Institute of Mental Health’s Research Domain Criteria (RDoC) (Cuthbert and Insel, 2013; Insel et al., 2010) treats cognition as one of the main basic functional domains that transdiagnostically underlie mental health. According to RDoC, mental health should be studied in relation to cognition, alongside other domains, such as negative and positive valence systems, arousal and regulatory systems, social processes, and sensorimotor functions. RDoC further emphasises that each domain, including cognition, should be investigated not only at the behavioural level but also through its neurobiological correlates. In this study, we aim to examine how the covariation between cognition and mental health is reflected in neural markers of cognition, as measured through multimodal neuroimaging. Recent efforts in brain Magnetic Resonance Imaging (MRI) and machine learning have led to predictive models that allow us to create MRI-based neural indicators of cognition with reasonable predictive performance (Krämer et al., 2024; Pat et al., 2022; Tetereva et al., 2022). These models are designed to predict cognition based on different cognitive tasks in unseen individuals who are not part of the modeling process (Marek et al., 2022; Zhi et al., 2024). Yet, the extent to which MRI-based neural indicators designed to predict cognition capture the same variance that mental health shares with cognition remains unknown. Demonstrating that MRI-based neural indicators of cognition capture the covariation between cognition and mental health will thereby support the utility of such indicators for understanding the etiology of mental health (Wang et al., 2025). Different MRI modalities measure different aspects of the brain, and MRI quantification techniques capture different brain features, resulting in distinct neuroimaging phenotypes. This means there are numerous approaches to derive neural indicators of cognition from MRI data. For example, diffusion-weighted MRI (dwMRI) measures the shape and amount of water diffusion in various directions and tissue compartments (Alexander et al., 2007). Different dwMRI metrics, such as fractional anisotropy (FA), which quantifies the degree of water diffusion directionality, and the streamline count, which indirectly reflects structural connectedness between the two regions (structural connectome), provide information about white matter orientation, density, and microstructural integrity (Basser et al., 1994; Soares et al., 2013; Zhang et al., 2012). Resting-state functional MRI (rsMRI) measures spontaneous low-frequency fluctuations in the Blood Oxygenation Level Dependent (BOLD) signal in the absence of a task, enabling the investigation of resting-state functional connectivity (RSFC) (Lee et al., 2013). RSFC from rsMRI can be estimated between pairs of parcellated grey matter regions (functional connectome) or between widespread networks derived from the Independent Component Analysis (ICA). Structural MRI (sMRI) uses T1-weighted and T2-weighted imaging to quantify various aspects of brain anatomy and morphology. For example, the morphology of the cerebral cortex and white matter can be quantified by measuring grey or white matter thickness, volume, and area in regions defined by different atlases, whereas the characteristics of subcortical regions are conventionally quantified with volumes of subcortical nuclei and their subdivisions (Symms et al., 2004; Wattjes, 2011). Previous studies using machine learning have shown that both (a) the choice of MRI modality and (b) the quantification method within each modality affect the performance of MRI-based models in capturing cognition (Dhamala et al., 2021; Pat et al., 2022; Tetereva et al., 2022). Dhamala and colleagues found that the predictive ability of structural and functional connectomes largely depends on the choice of atlases used to parcellate grey matter and how they were derived (Dhamala et al., 2021). Given the heterogeneity of neuroimaging phenotypes from different MRI modalities, drawing information across them may boost the predictive ability of MRI-based neural indicators (Caunca et al., 2021). One way to integrate multiple neuroimaging phenotypes across MRI modalities is a stacking approach, which employs two levels of machine learning. First, researchers build a predictive model from each neuroimaging phenotype (e.g. cortical thickness from different grey matter parcellations) to predict a target variable (e.g. cognition). Next, in the stacking level, they use predicted values (i.e. cognition predicted from each neuroimaging phenotype) from the first level as features to predict the target variable (Pat et al., 2022). Previous studies show that integrating multimodal neuroimaging phenotypes into ‘stacked models’ enhances the prediction of cognition (Krämer et al., 2024; Pat et al., 2022; Rasero et al., 2021; Tetereva et al., 2022). Here, we aim to determine whether this improvement extends beyond the prediction of cognition itself, allowing us to capture more covariation between cognition and mental health. Using the largest population-level neuroimaging dataset, the UK Biobank, we investigated (a) which neuroimaging phenotypes yield a neural indicator of cognition that explains the relationship between cognition and mental health the most, and (b) whether combining neuroimaging phenotypes within and across MRI modalities enhances the explanation of this relationship. We started by deriving a general cognition factor, or the g-factor, from twelve cognitive scores from different tasks. The g-factor underlies variability across cognitive domains and reflects the overall cognition (Jensen, 2000; Panizzon et al., 2014). Next, we applied machine learning to predict the g-factor from 133 mental health indices and 72 neuroimaging phenotypes in unseen participants. For neuroimaging, we created predictive models from both individual neuroimaging phenotypes and phenotypes combined within and across three MRI modalities via stacking. Finally, we conducted commonality analyses (Nimon et al., 2008) to quantify the contribution of neural indicators of cognition based on different neuroimaging phenotypes to explaining the relationship between cognition and mental health. Results g-factor-modeling To model the g-factor, we split the data into five outer folds, each comprising training (80% of the data) and test (20% of the data) sets (see Table 1 for sample characteristics and the Data analysis section). In each fold, data factorability, assessed using Kaiser-Meyer-Olkin statistics (KMO >0.87) and Bartlett’s test of sphericity (p 0.96, the Tucker-Lewis Index (TLI) > 0.92, the Root Mean Square Error of Approximation (RMSEA) ≤ 0.05, and the Standardized Root Mean Square Residual (SRMR) 0.10 poor fit. The SRMR measures the average standardized difference between observed and model-implied correlations and is less sensitive to variable scaling than the Root Mean Square Residual; values 26 indicating very high risk (Davis et al., 2020; Weathers et al., 2018). PTSD status was coded as positive for PCL-6 scores ≥14 and included stressful life events in addition to catastrophic trauma. Alcohol use, dependence, and alcohol-related harm were assessed using the AUDIT questionnaire, calculated as the sum of ten items (Saunders et al., 1993). The AUDIT score was further partitioned into alcohol consumption (AUDIT-C; items 1–3) and alcohol-related problems (AUDIT-P; items 4–10) (Sanchez-Roige et al., 2019). For items 2–10, missing responses were replaced with zero when the first screening question (‘Frequency of drinking alcohol’) indicated no alcohol consumption, as these responses correspond to ‘Never.’ An AUDIT score of 8 was used as a threshold for hazardous alcohol use, with scores of 8–15 indicating harmful drinking and scores >15 indicating alcohol dependence or moderate-to-severe alcohol use disorder (Babor et al., 2001; Saunders et al., 1993). Hazardous alcohol use and alcohol dependence were, therefore, defined as AUDIT scores ≥8 and ≥15, respectively. ‘Alcohol dependence ever’ status was coded as positive if a participant reported a history of physical dependence on alcohol. To reduce skewness, AUDIT, AUDIT-C, and AUDIT-P scores were log(x+1)-transformed (Sanchez-Roige et al., 2019). Brain MRI MRI data were collected on a Siemens Magnetom Skyra 3T scanner with the Siemens 32-channel head coil across four imaging centres: Cheadle, Reading, Newcastle, and Bristol. The data were processed in FSL (FMRIB Software Library). For MRI-based models, we used dwMRI, rsMRI, and sMRI (Figure 1 and Supplementary file 4). MRI acquisition protocols are available at http://biobank.ctsu.ox.ac.uk/crystal/refer.cgi?id=2367. MRI processing pipelines are described in the UK Biobank brain imaging documentation (https://biobank.ctsu.ox.ac.uk/crystal/crystal/docs/brain_mri.pdf) and discussed elsewhere (Alfaro-Almagro et al., 2018; Mansour et al., 2023). MRI data processed by the UK Biobank are available as imaging-derived phenotypes (IDPs). In addition to the UK Biobank’s IDPs, we used structural connectomes and BOLD time series created by Mansour and colleagues (Mansour et al., 2023). Diffusion-weighted MRI (dwMRI) Request a detailed protocoldwMRI data were acquired using a spin-echo echo-planar imaging sequence with 3x multislice (multiband) acceleration in 100 diffusion-encoding directions across two shells: 50 directions with b=1000 s/mm² and 50 directions with b=2000 s/mm² (TE = 92 ms, TR = 3600 ms, 2 mm isotropic voxels, partial Fourier 6/8, FOV: 104×104×72, δ=21.4 ms) in two phase-encoding directions. Primary data were collected with anterior-to-posterior phase encoding, and three b=0 images with reversed phase encoding were acquired for field map estimation and distortion correction. To quantify dwMRI, we used 42 IDPs. These IDPs were obtained from head motion-, eddy current-, and gradient distortion-corrected dwMRI data and included fractional anisotropy (FA), diffusion tensor mode (MO), mean diffusivity (MD), and eigenvalues of the diffusion tensor (L1, L2, L3) derived from diffusion tensor fitting, as well as intracellular volume fraction (ICVF), isotropic/free water volume fraction (ISOVF), and orientation dispersion index (OD) from neurite orientation dispersion and density imaging (NODDI). Diffusion metrics were computed for 48 and 27 white matter tracts reconstructed using tract-based spatial statistics (TBSS) and probabilistic tractography, respectively (Andersson and Sotiropoulos, 2016; Behrens et al., 2003; Behrens et al., 2007; Glasser et al., 2013; Jbabdi et al., 2012; Smith et al., 2006). FA values range from 0 to 1 and reflect the degree of directional water diffusion (anisotropy). MD represents the directionally averaged diffusion rate, independent of orientation (Alexander et al., 2007; O’Donnell and Westin, 2011; Soares et al., 2013). MO is perpendicular to FA. It quantifies tensor shape and distinguishes between linear (tube-like) and planar (disk-like) diffusion patterns (Tae et al., 2018; Yoncheva et al., 2016). L1, L2, and L3 represent diffusion along the three principal axes of the diffusion tensor (O’Donnell and Westin, 2011). NODDI-derived ICVF measures water diffusion within neurites, reflecting neurite density; ISOVF represents freely diffusing water (e.g. cerebrospinal fluid); and OD measures neurite dispersion (spatial configuration), with higher values typically observed in gray matter and lower values in aligned white matter tracts (Colgan et al., 2016; Deligianni et al., 2016; Zhang et al., 2012). Probabilistic tractography was performed in FSL using BEDPOSTx and PROBTRACKx, and diffusion metrics were mapped onto 27 fibre tracts using standard-space region-of-interest masks defined by AutoPtx (de Groot et al., 2013). In probabilistic tractography, white matter tracts are reconstructed by estimating probability distributions of voxel-to-voxel connections to account for uncertainty in fibre orientation (Morris et al., 2008). FA images derived from diffusion tensor fitting were also processed using TBSS (Smith et al., 2006). In TBSS, individual FA images are projected onto a standard-space mean FA skeleton representing fibre tracts common across participants, and the resulting spatial transformation is applied to other microstructural indices to enable between-subject comparisons (Smith et al., 2006). Skeletonized FA maps and DTI/NODDI metrics were averaged across 48 standard-space tract masks from the Susumu Mori white matter atlas (Mori et al., 2006; Wakana et al., 2007). In addition to tract-based metrics, we used structural connectomes derived from six combinations of cortical and subcortical atlases: Desikan-Killiany (aparc)+Melbourne Subcortical Atlas scale I (MSA-I), Destrieux (aparc.a2009s)+MSA I, Glasser +MSA I, Glasser+MSA IV, Schaefer-200 (7 networks)+MSA I, and Schaefer-500 (7 networks)+MSA IV (Desikan et al., 2006; Destrieux et al., 2010; Glasser et al., 2016; Schaefer et al., 2018; Thomas Yeo et al., 2011; Tian et al., 2020). Structural connectome data consisted of connectivity matrices for each atlas combination containing streamline count, fibre bundle capacity (SIFT2), mean streamline length, and mean FA for each pair of nodes (Supplementary file 4; Mansour et al., 2023). A procedure for computing structural and functional connectomes is discussed in detail in Mansour et al., 2023. Briefly, the connectomes were generated following the Anatomically Constrained Tractography (ACT) framework using fibre orientation distributions (iFOD2) probabilistic fibre tracking. In this approach, streamlines are propagated from seed points within white matter or at the gray-white matter interface based on local fibre orientation distributions or dynamic seeding using spherical-deconvolution informed filtering of tractograms (SIFT/SIFT2) (Smith et al., 2012; Smith et al., 2013a; Smith et al., 2015). SIFT and SIFT2 correct for biases in streamline density by aligning tractogram density with the underlying fibre density estimated from the diffusion signal, thereby improving the biological accuracy of reconstructed pathways (Smith et al., 2013b; Smith et al., 2015). SIFT2 assigns weights to streamlines, and the sum of these weights represents fibre bundle capacity, corresponding to the estimated bundle’s total intracellular cross-sectional area and connectivity strength of white matter pathways (Smith et al., 2020). In preprocessing, averaged b=0 volumes were skull-stripped using FSL BET to improve T1-weighted and dwMRI registration. Tissue response functions for white matter, gray matter, and cerebrospinal fluid were used to estimate fibre orientation distributions and perform intensity normalization and bias field correction. Whole-brain probabilistic tractography was then conducted within the ACT framework using tissue-type segmentation obtained from FreeSurfer and FSL FIRST, with streamline seeding initiated at the gray-white matter interface, resulting in 10 million streamlines satisfying ACT constraints (Mansour et al., 2023; Smith et al., 2012). Whole-brain tractograms were subsequently parcellated using cortical and subcortical atlases defined in template space and transformed into native space. Subcortical parcellations were generated by applying the inverse native T1w-to-MNI152 warp using FSL applywarp, while cortical labels were aligned to individual cortical surfaces using FreeSurfer fsnative (Mansour et al., 2023). Cortical and subcortical parcellations were combined to construct structural connectomes, enabling calculation of streamline-based connectivity metrics between all gray matter regions. Resting-state functional MRI (rsMRI) Request a detailed protocolrsMRI data were acquired using a gradient-echo echo-planar imaging sequence with 8x multislice (multiband) acceleration (TE = 39 ms, TR = 735 ms, 2.4 mm isotropic voxels, 490 volumes, FOV: 88×88×64, flip angle = 52°). Preprocessing followed the UK Biobank pipeline and included correction for susceptibility- and gradient-induced distortions, motion correction using FMRIB’s Linear Image Registration Tool (MCFLIRT), grand-mean intensity scaling, high-pass temporal filtering, and structured artefact removal using independent component analysis (ICA) combined with FMRIB’s ICA-based X-noiseifier (FIX), with non-neural components discarded. Group ICA was performed using FSL MELODIC to derive spatial maps of resting-state networks, followed by dual regression to estimate subject-specific BOLD time series for each network (Beckmann et al., 2009; Beckmann, 2012; Beckmann and Smith, 2004; Nickerson et al., 2017). To quantify rsMRI, we used 26 IDPs. These included functional connectivity measures – normalized full and partial temporal correlations between node time series – and network amplitudes (standard deviation of network time series) for 21 and 55 independent components (ICs), representing statistically independent or spatially distinct resting-state functional networks, obtained from ICA decompositions at 25 and 100 dimensions, respectively (Beckmann et al., 2005; Cole et al., 2010; Lee et al., 2023). Partial correlations were regularized using ridge regression (L2 regularization, ρ=0.5) to improve estimation stability. All correlation coefficients were Gaussianized (z-transformed) and corrected for temporal autocorrelation. In addition to ICA-based metrics, we derived functional connectomes from parcellated BOLD time series using six combinations of cortical and subcortical atlases, consistent with the structural connectome analysis (Mansour et al., 2023). Cortical and subcortical time series, provided separately, were combined prior to connectivity estimation. Parcellation-based functional connectomes were constructed by averaging preprocessed BOLD time series within regions of interest. Atlas labels were resampled from the individual T1-weighted space to the fMRI voxel grid using FreeSurfer’s mri_vol2vol, and voxel-wise BOLD signals were averaged within each region. Functional connectivity matrices were then computed using the ConnectivityMeasure function from the Nilearn Python library. For both ICA- and atlas-based approaches, we complemented correlation-based metrics with tangent space parameterization, defined as deviations of individual connectivity matrices from the group-average connectivity. This involved estimating a group-average covariance matrix from the training data using the Ledoit-Wolf shrinkage estimator and projecting individual covariance matrices from both training and test sets into a common tangent space (Abbas et al., 2023; Dadi et al., 2019; Ledoit and Wolf, 2004; Ng et al., 2014; Pervaiz et al., 2020). Prior to projection, BOLD time series were z-score standardized. The resulting matrices were vectorized, and diagonal elements were excluded. This approach accounts for the non-Euclidean geometry of covariance and correlation matrices, which lie on a high-dimensional nonlinear curved surface – Riemannian manifold of symmetric positive definite matrices (Abbas et al., 2023; Venkatesh et al., 2020). Projection into a common tangent space reduces statistical dependencies between covariance estimates and enables the use of conventional linear methods, addressing limitations related to the nonlinear geometry of functional connectivity data and often improving performance relative to standard correlation-based measures (Abbas et al., 2023; Dadi et al., 2019; Ng et al., 2014; Simeon et al., 2022; Venkatesh et al., 2020). For each atlas combination, full and partial correlations were computed across the entire sample, while tangent space representations were fitted on the training set and applied to both training and test sets within each cross-validation fold. Pearson correlation coefficients were additionally transformed using the inverse hyperbolic tangent (arctanh) function to reduce skewness. Machine learning models were trained separately for each connectivity measure, and the measure with the best training performance was selected for downstream analyses. Structural MRI (sMRI) Request a detailed protocolT1w images were acquired using a 3D MPRAGE sequence (1 mm isotropic voxels, FOV: 208×256×256). Structural preprocessing included gradient distortion correction, brain extraction and defacing, and registration from native space to MNI152 standard space using linear and nonlinear transformations (Jenkinson et al., 2002; Jenkinson et al., 2012; Jenkinson and Smith, 2001; Mansour et al., 2023; Smith, 2002). Tissue-specific and total brain volumes were estimated using SIENAX-style analysis (Structural Image Evaluation, using Normalization, of Atrophy: Cross-sectional) (Smith et al., 2002). Shapes and volumes of 15 subcortical structures were modeled using FIRST (FMRIB’s Integrated Registration and Segmentation Tool) (Patenaude et al., 2011). Cortical surface area and thickness were estimated with FreeSurfer, and subcortical segmentations were obtained using the ASEG tool (Fischl, 2012). T2w images were acquired using a 3D SPACE sequence with fluid-attenuated inversion recovery (FLAIR) contrast to enhance visualization of white matter hyperintensities (1.05×1×1 mm voxels, FOV: 192×256×256). T2w images were linearly registered to T1w images using FSL FLIRT (FMRIB’s Linear Image Registration Tool), subsequently aligned to MNI space, and bias-field corrected. Total white matter hyperintensity volume (i.e. white matter lesions) was quantified using the BIANCA (Brain Intensity AbNormality Classification Algorithm) (Griffanti et al., 2016). Tissue segmentation and bias-field correction were performed in FSL using FAST (FMRIB’s Automated Segmentation Tool) (Zhang et al., 2001). To quantify sMRI, we used 20 IDPs, which included gray and white matter volumes and mean intensities, cortical surface area and thickness, volumes of subcortical structures and cerebrospinal fluid, gray-white matter contrast (fractional intensity contrast), and total brain volume. These measures were derived using eight segmentation strategies: FSL FAST regional gray matter volumes (Zhang et al., 2001), FSL FIRST subcortical volumes (Patenaude et al., 2011), FreeSurfer ASEG subcortical segmentation, FreeSurfer ex-vivo Brodmann area maps, FreeSurfer Destrieux (a2009s) atlas, FreeSurfer Desikan-Killiany-Tourville parcellation, FreeSurfer Desikan-Killiany parcellation (including gray/white matter intensity and pial and white matter surfaces), and FreeSurfer subcortical subsegmentation (Iglesias et al., 2015). T2w-derived IDPs included total volumes of deep white matter, periventricular white matter, and white matter hyperintensities (Supplementary file 4). MRI confounds Request a detailed protocolWe adjusted all MRI modalities for common and modality-specific confounds as described by Alfaro-Almagro et al., 2021. Common confounds included scanning site, acquisition date, head size, scanner table position, radiofrequency receive coil position, brain position within the scanner, mismatch between T1w images and the standard-space template, and structural motion (head motion estimated from T1w images) (Alfaro-Almagro et al., 2021). Head size was defined as a volumetric scaling factor obtained during transformation from native T1w space to MNI space using SIENAX (Alfaro-Almagro et al., 2018; Smith et al., 2002). Scanning site (four UK Biobank centres) was encoded using dummy variables, and the acquisition date was converted to Unix time. Modality-specific confounds included head motion (mean displacement between consecutive time points, averaged across the brain and time), intensity scaling, and mismatch between non-structural images and the T1w reference. Motion summaries additionally included median absolute and relative displacement (Supplementary file 4). For dwMRI, motion estimates were obtained using FSL eddy. To account for slow signal drifts (e.g. heating effects), we included additional confounds related to advanced eddy current correction, such as expanded search space parameters. We also included the number of slices identified as outliers by eddy as a motion-related confound. For rsMRI, motion parameters were estimated using FSL FEAT (FMRI Expert Analysis Tool). Additional motion-related confounds were derived using DVARS, which decomposes signal variability into fast (D-var), slow (S-var), and edge (E-var) components (Afyouni and Nichols, 2018). D-var represents the mean squared difference between consecutive volumes, S-var reflects the mean squared average of adjacent volumes, and E-var captures edge-related variability at the beginning and end of the time series. DVARS-based confounds included the mean, median, and 90th percentile of S-var and D-var, normalized by total signal variance (A-var). All confounds were regressed out from MRI features following standardization of both confounds and features. Interactive visualizations of variance explained by confound groups are available at https://www.fmrib.ox.ac.uk/ukbiobank/confounds/plots_2020_03_11/index.html. Data analysis Machine learning Request a detailed protocolTo model the relationship between mental health and cognition, we employed PLSR to predict the g-factor from 133 mental health variables, and to model the relationship between brain and cognition, we used a two-step stacking approach (Krämer et al., 2024; Pat et al., 2022; Tetereva et al., 2022; Tetereva et al., 2025) that integrates information from neuroimaging phenotypes across three MRI modalities. In the first step, we trained 72 base (first-level) PLSR models, each predicting the g-factor from a single neuroimaging phenotype. In the second step, we used the predicted values from these base models as input features for stacked models, which again predicted the g-factor. We constructed four stacked models based on the source of the base predictions: one each for dwMRI (‘dwMRI Stacked’), rsMRI (‘rsMRI Stacked’), sMRI (‘sMRI Stacked’), and a combined model incorporating all modalities (‘All MRI Stacked’). Each stacked model was trained using one of four machine learning algorithms – ElasticNet, Random Forest, XGBoost, or Support Vector Regression (SVR) – selected individually for each model. Analyses were performed in Python (version 3.11), and machine learning models were implemented using the scikit-learn library. We employed nested cross-validation to predict cognition from mental health indices and neuroimaging phenotypes (Figure 1). Nested cross-validation is a robust method for evaluating machine learning models while tuning their hyperparameters, ensuring that performance estimates are both accurate and unbiased. Here, we used a nested cross-validation scheme with five outer folds and ten inner folds (Wilimitis and Walsh, 2023). We started by dividing the entire dataset into five outer folds. Each fold took a turn being held out as the outer-fold test set (20% of the data), while the remaining fourfolds (80% of the data) were used as an outer-fold training set. Within each outer-fold training set, we performed a second layer of cross-validation – this time splitting the data into ten inner folds. These inner folds were used exclusively for hyperparameter tuning: models were trained on nine of the inner folds and validated on the remaining one, cycling through all ten combinations. We then selected the hyperparameter configuration that minimized the mean squared error (MSE) across inner-fold validation sets. The model was then retrained on the full outer-fold training set using the optimal configuration and evaluated on the outer-fold test set using Pearson correlation (r), the coefficient of determination (R2), mean absolute error (MAE), and MSE for the predicted and observed g-factor values. The evaluation metrics were computed as follows: where yi, i, and ŷi denote observed values, the mean of observed values, and predicted values, respectively. Performance metrics were averaged across the five outer folds for each feature set. For rsMRI IDPs, the choice of functional connectivity quantification method – full correlation, partial correlation, or tangent space parameterization – was treated as a hyperparameter. The method yielding the best performance in the inner cross-validation (training) folds was selected and subsequently evaluated on the outer-fold test sets for predicting the g-factor. This entire process was repeated for each of the five outer folds, ensuring that every data point is used for both training and testing, but never at the same time. We opted for five outer folds instead of ten to reduce computational demands, particularly memory and processing time, given the substantial volume of neuroimaging data involved in model training. Five outer folds led to an outer-fold test set at least n=4000, which should be sufficient for model evaluation. In contrast, we retained ten inner folds to ensure robust and stable hyperparameter tuning, maximising the reliability of model selection. The first-level machine learning model applied to mental health and MRI data was PLSR. PLSR is a multivariate method that simultaneously performs dimensionality reduction and regression by projecting correlated predictor variables onto a set of latent components that maximise covariance with the response variable(s) (Wold et al., 2001). This makes it particularly suitable for high-dimensional and collinear neuroimaging and behavioural data. Formally, PLSR decomposes the predictor matrix X into orthogonal latent scores and loadings and models the response variable Y as a linear function of these components, yielding a regression model of the form: where B represents regression coefficients and F residuals. This is achieved by projecting the original predictors onto a set of latent variables (X-scores), which are weighted linear combinations of the original variables: where T is the matrix of latent scores, P the loadings, W* the transformed (rotated) weight matrix, and E the residual matrix. The scores T represent projections of the observations onto the latent components, whereas the loadings P describe how the original variables contribute to these components and are represented in the latent space. The weights determine how the original variables are combined to form the latent scores. Specifically, weights are estimated by maximising the covariance between X and Y, thereby capturing the most predictive variation in the data, and are subsequently normalized so that the latent scores have unit variance. The response variable is then modeled as a function of the latent scores: where C is the matrix of Y-weights (with columns ca) that relate the latent components to the response variable(s). In the case of multiple response variables, PLSR similarly defines Y-scores U=YC, such that Y=UC′+G, where U represents projections of Y onto the latent space. This leads to the regression formulation: where B=W*C′. In this framework, the latent components act as orthogonal predictors, reducing multicollinearity while preserving information relevant for predicting Y. In practice, PLSR iteratively extracts components, with each step followed by deflation of the predictor matrix to ensure that subsequent components capture new, orthogonal sources of covariance between X and Y (Wold et al., 2001). This deflation procedure ensures that the extracted scores are orthogonal, while weights remain orthonormal across components. Thus, the latent variables provide a low-dimensional representation of the predictors optimized for the prediction of the response. Geometrically, X-scores (ta, i.e. the score vectors corresponding to the columns of the score matrix T) can be interpreted as coordinates of the projection of the predictor matrix X onto an A-dimensional latent hyperplane that approximates the original data. Each component defines a direction within this space, with X-loadings (pa, i.e. the loading vectors corresponding to the columns of the loading matrix P) representing the corresponding direction coefficients (slopes) that describe how the original variables contribute to these components. These directions are chosen to maximise the covariance between X and Y, thereby defining the latent space that best captures their shared structure. A comprehensive overview of the mathematical and geometric interpretation of PLSR is available at: https://learnche.org/pid/latent-variable-modelling/projection-to-latent-structures/index. In summary, PLSR constructs new predictor variables (X-scores) that (i) represent the original predictors in a latent space, (ii) are linear combinations of the original variables, and (iii) are maximally predictive of the response variable. The only hyperparameter in PLSR is the number of components. We optimized this parameter using nested cross-validation with random shuffling. The optimal number of components was selected via grid search (GridSearchCV, scikit-learn) by minimising the MAE. The final model was retrained on the full outer-fold training set using the optimal number of components and evaluated on the held-out test set to obtain out-of-sample predictions of the g-factor. To train the second-level models, we evaluated four machine learning algorithms – ElasticNet, Random Forest, XGBoost, and SVR – and selected the model with the highest out-of-sample R2 (Chicco et al., 2021). The hyperparameter grids for all models are provided in Supplementary file 5. ElasticNet Request a detailed protocolElasticNet is a linear regression model that combines L1 (Lasso) and L2 (Ridge) regularization to address multicollinearity and improve generalization (Hoerl, 2020; Kotz et al., 1982; Tibshirani, 1996; Zou and Hastie, 2005). The L1 penalty promotes sparsity by shrinking some coefficients to zero (enabling variable selection), whereas the L2 penalty stabilises estimates by shrinking coefficients toward zero without eliminating them. The mixing parameter (l1_ratio) controls the relative contribution of L1 and L2 penalties (0 = Ridge, 1 = Lasso), and the regularization strength (alpha) controls the overall degree of shrinkage. This combination allows ElasticNet to retain groups of correlated predictors while reducing overfitting. Thus, ElasticNet addresses several limitations of Lasso regression: it can handle settings with more predictors than observations; it avoids selecting only a single variable from a group of correlated predictors; and it improves prediction accuracy in the presence of multicollinearity (van Erp et al., 2019). Furthermore, by combining L1 and L2 regularization, ElasticNet promotes more parsimonious models while mitigating the redundancy characteristic of Ridge regression, which tends to retain all correlated predictors through continuous shrinkage (Zou and Hastie, 2005). Support vector regression (SVR) Request a detailed protocolSVR extends Support Vector Machines (SVM), a supervised learning method introduced by Vapnik, to regression problems (Awad and Khanna, 2015; López et al., 2022; Rodríguez-Pérez and Bajorath, 2022; Vapnik, 1998; Vapnik, 1995). SVMs can capture nonlinear relationships and handle high-dimensional data, including settings with many predictors and relatively few observations, by implicitly mapping the original data into a higher-dimensional feature space using the kernel trick. This approach computes inner products between observations in the transformed space without explicitly performing the transformation, enabling efficient modeling of complex relationships. In the SVM framework, the goal is to identify an optimal hyperplane that maximises the margin between data points. The margin is defined by the distance between the hyperplane and the closest data points, known as support vectors (SVs), which determine the position and orientation of the decision boundary. In SVR, this concept is adapted to regression through the introduction of an ε-insensitive tube around the regression function. Errors within this tube are not penalized, whereas deviations exceeding ε (the distance from boundaries to the hyperplane) are penalized via an ε-insensitive loss function. The model, therefore, seeks a function that is both as flat as possible and captures the majority of the data within this tolerance region. Support vectors in SVR are the observations that lie outside the ε-tube and thus define the shape and position of the regression function. To capture nonlinear relationships, we used the radial basis function (RBF) kernel, which measures similarity between observations based on their distance in the input space. In other words, in the regression setting, the RBF kernel enables the estimation of similarity between new observations and support vectors, which define the shape and position of the regression function (Elish, 2014; Ramedani et al., 2014; Smola and Schölkopf, 2004). It maps the data into a higher-dimensional feature space where a linear relationship can approximate complex nonlinear patterns in the original space (Chen and Bakshi, 2009; Ramedani et al., 2014). This allows SVR to model nonlinear dependencies by effectively ‘linearising’ them after transformation. The kernel width parameter (γ) controls how rapidly this similarity decays with distance, thereby determining the strength of the influence of a single training instance. Larger γ values make the model focus more strongly on nearby observations (i.e. highly local influence), resulting in more flexible but potentially overfitted solutions, whereas smaller γ values lead to broader, smoother patterns that generalise better. The regularization parameter (C) controls the trade-off between fitting the training data and limiting model complexity. Higher values of C prioritisze minimizing prediction errors, allowing the model to fit the training data more closely, whereas lower values tolerate more errors but yield simpler models that generalize better to unseen data (Awad and Khanna, 2015; Balfer and Bajorath, 2015; Liu and Zheng, 2006; Rodríguez-Pérez and Bajorath, 2022). In practical terms, C determines how strongly the model penalises deviations outside the ε-insensitive tube. Random forest Request a detailed protocolRandom Forest is an ensemble learning method within the Classification and Regression Tree (CART) framework that constructs multiple decision trees on bootstrap samples of the data and aggregates their predictions (Breiman, 2001; Breiman et al., 2017; Schonlau and Zou, 2020; Svetnik et al., 2003). It employs two sources of randomness – bootstrap sampling of observations and random selection of predictor subsets at each split (max_features, also referred to as mtry) – to decorrelate trees, reduce variance, and improve generalization relative to a single decision tree. Each tree is typically grown to substantial depth (constrained by max_depth and minimum node size), allowing the model to capture complex nonlinear relationships and interactions. Predictions are obtained by averaging outputs across trees, which stabilises estimates and smooths nonlinearities. Splits within each tree are selected to maximize the reduction in impurity, ensuring greater homogeneity of the resulting nodes. In regression settings, impurity is measured by the residual sum of squares, and splits are chosen to minimize within-node variance. The magnitude of impurity reduction also provides a basis for estimating feature importance, with larger reductions indicating greater predictive contribution (Breiman et al., 2017; Ishwaran, 2015; Nembrini et al., 2018). Key hyperparameters include the number of decision trees built on different bootstrap samples (n_estimators), which controls the size of the ensemble; the maximum tree depth (max_depth), which determines the number of splits within each tree, governing model complexity; and the number of predictors considered at each split (max_features). XGBoost Request a detailed protocolXGBoost (eXtreme Gradient Boosting) is a gradient boosting framework that sequentially builds decision trees to minimise a specified loss function, i.e., the discrepancy between observed and predicted values of the response variable, using gradient descent (Chen and Guestrin, 2016; Tarwidi et al., 2023). Each new tree is fitted to the residuals of the previous ensemble, progressively improving model performance. In other words, the model iteratively corrects errors made by earlier trees. XGBoost incorporates regularization of tree complexity and feature weights to reduce overfitting and enhance generalizability. Standardization and statistical significance Request a detailed protocolTo prevent data leakage, we standardized the data using the mean and standard deviation derived from the training set and applied these parameters to the corresponding test set within each outer fold. This standardization was performed at three key stages: before g-factor derivation, before regressing out modality-specific confounds from the MRI data, and before stacking. Similarly, to maintain strict separation between training and testing data, both base and stacked models were trained exclusively on participants from the outer-fold training set and subsequently applied to the corresponding outer-fold test set. To assess the statistical significance of the models, we aggregated the predicted and observed g-factor values from each outer-fold test set. We then computed a bootstrap distribution of Pearson’s correlation coefficient (r) by resampling with replacement 5000 times, generating 95% confidence intervals (CIs). Model performance was considered statistically significant if the 95% CI did not include zero, indicating that the observed associations were unlikely to have occurred by chance. Feature importance Request a detailed protocolTo identify the neuroimaging features that contribute most to the predictive performance of top-performing phenotypes within each modality, while accounting for the potential latent components derived from neuroimaging, we assessed feature importance using the Haufe transformation (Haufe et al., 2014). Specifically, we calculated Pearson correlations between the predicted g-factor and scaled and centred neuroimaging features across five outer-fold test sets. We also examined whether the performance of neuroimaging phenotypes in predicting cognition per se is related to their ability to explain the link between cognition and mental health. Here, we computed the correlation between the predictive performance of each neuroimaging phenotype and the proportion of the cognition–mental health relationship it captures. Commonality analysis Request a detailed protocolFinally, we conducted a series of commonality analyses to quantify the proportion of the cognition–mental health relationship captured by MRI-based neural indicators (Nimon et al., 2008; Seibold and McPHEE, 1979; Viswesvaran, 1998). For each set of MRI data (i.e. individual neuroimaging phenotypes, neuroimaging phenotypes stacked within each MRI modality, and neuroimaging phenotypes stacked across all MRI modalities), for both observed (i.e. derived directly from cognitive scores, not mental health or MRI data) and predicted g-factors (i.e. predicted from either mental health indices or neuroimaging phenotypes using the top-performing algorithm), we separately pooled values across all five outer-fold test sets to reconstruct the complete dataset. We then applied a series of linear regression models, treating the observed g-factor as a response variable and the g-factor predicted from mental health (mental health-g) and neuroimaging phenotypes (neuroimaging-g) as two separate explanatory variables: Next, we decomposed the total variance of the observed g-factor, R2, into the variance explained uniquely or commonly by mental health-g and neuroimaging-g, as follows Nimon et al., 2008; Seibold and McPHEE, 1979; Wang et al., 2025: We quantified the contribution of each neuroimaging phenotype to explaining the cognition–mental health relationship by a percentage ratio between (a) the common effects between mental health-g and neuroimaging-g and (b) the total effects of mental health-g. That is, the numerator is the common variance that neuroimaging-g shares with mental health-g in explaining the observed g-factor, and the denominator is the variance of mental health-g that explains the observed g-factor regardless of neuroimaging-g: where R2mental health = Commonmental health, neuroimaging + Uniquemental health. Each commonality analysis was conducted independently and, therefore, included a different number of participants, depending on how many were available for a given neuroimaging phenotype or modality. To understand how demographic factors, including age and sex, contribute to this relationship, we also conducted a separate set of commonality analyses treating age, sex, age2, age×sex, and age2×sex as an additional set of explanatory variables. Appendix 1 Data availability This study used data from the UK Biobank resource (Application No. 70132). These data cannot be publicly shared by the authors due to legal and ethical restrictions imposed by UK Biobank. Access to individual-level data is governed by UK Biobank's data access policies, which are designed to protect participant confidentiality. Researchers can access the data by submitting an application directly to UK Biobank (https://www.ukbiobank.ac.uk/enable-your-research/apply-for-access). Applications must include a research proposal outlining the scientific rationale and intended use of the data, and are reviewed as part of UK Biobank's access procedures. Access may be granted to approved researchers from academic, charity, government, and commercial organisations, subject to UK Biobank's terms and conditions. Due to these restrictions, we are not permitted to share raw or deidentified individual-level data. Redistribution of such data is contractually prohibited, and deidentification does not eliminate the risk of participant re-identification. This manuscript is a computational study and did not generate new primary data. Numerical data underlying the figures are provided as figure source data files. All analyses can be reproduced using the code provided by the authors. The modelling and analysis code is openly available on GitHub: https://github.com/HAM-lab-Otago-University/UKBiobank/ (copy archived at Buianova, 2026). References - The C factor: cognitive dysfunction as a transdiagnostic dimension in psychopathologyClinical Psychology Review 86:102007.https://doi.org/10.1016/j.cpr.2021.102007 - Insight and inference for DVARSNeuroImage 172:291–312.https://doi.org/10.1016/j.neuroimage.2017.12.098 - Diffusion tensor imaging of the brainNeurotherapeutics 4:316–329.https://doi.org/10.1016/j.nurt.2007.05.011 - Executive function and PTSD: disengaging from traumaNeuropharmacology 62:686–694.https://doi.org/10.1016/j.neuropharm.2011.02.008 - BookAUDIT-The Alcohol Use Disorders Identification Test: Guidelines for Use in Primary Heath CareWorld Health Organization. - A note on the multiplying factors for various χ2 approximationsJournal of the Royal Statistical Society Series B 16:296–298.https://doi.org/10.1111/j.2517-6161.1954.tb00174.x - MR diffusion tensor spectroscopy and imagingBiophysical Journal 66:259–267.https://doi.org/10.1016/S0006-3495(94)80775-1 - Social context of drinking among older adults: relationship to alcohol and traffic risk behaviorsTransportation Research. Part F, Traffic Psychology and Behaviour 64:161–170.https://doi.org/10.1016/j.trf.2019.05.001 - Probabilistic independent component analysis for functional magnetic resonance imagingIEEE Transactions on Medical Imaging 23:137–152.https://doi.org/10.1109/TMI.2003.822821 - Investigations into resting-state connectivity using independent component analysisPhilosophical Transactions of the Royal Society of London. Series B, Biological Sciences 360:1001–1013.https://doi.org/10.1098/rstb.2005.1634 - Characterization and propagation of uncertainty in diffusion-weighted MR imagingMagnetic Resonance in Medicine 50:1077–1088.https://doi.org/10.1002/mrm.10609 - Comparative fit indexes in structural modelsPsychological Bulletin 107:238–246.https://doi.org/10.1037/0033-2909.107.2.238 - The relationship between cognitive control and posttraumatic stress symptomsJournal of Behavior Therapy and Experimental Psychiatry 43:844–848.https://doi.org/10.1016/j.jbtep.2011.12.001 - White matter damage in Alzheimer’s disease assessed in vivo using diffusion tensor magnetic resonance imagingJournal of Neurology, Neurosurgery, and Psychiatry 72:742–746.https://doi.org/10.1136/jnnp.72.6.742 - BookClassification And Regression TreesChapman and Hall/CRC.https://doi.org/10.1201/9781315139470 - Alcohol and social connectedness for new residential university students: implications for alcohol harm reductionJournal of Further and Higher Education 44:216–230.https://doi.org/10.1080/0309877X.2018.1527024 - Screening for bipolar spectrum disorders: a comprehensive meta-analysis of accuracy studiesJournal of Affective Disorders 172:337–346.https://doi.org/10.1016/j.jad.2014.10.024 - The Adolescent Brain Cognitive Development (ABCD) study: imaging acquisition across 21 sitesDevelopmental Cognitive Neuroscience 32:43–54.https://doi.org/10.1016/j.dcn.2018.03.001 - The prevalence of bipolar disorder in general primary care samples: a systematic reviewGeneral Hospital Psychiatry 36:19–25.https://doi.org/10.1016/j.genhosppsych.2013.09.008 - Book3.12—linear approaches for nonlinear modelingIn: Brown SD, Tauler R, Walczak B, editors. Comprehensive Chemometrics. Elsevier. pp. 453–462.https://doi.org/10.1016/B978-044452701-1.00060-0 - Advances and pitfalls in the analysis and interpretation of resting-state FMRI dataFrontiers in Systems Neuroscience 4:8.https://doi.org/10.3389/fnsys.2010.00008 - Multi-task connectivity reveals flexible hubs for adaptive task controlNature Neuroscience 16:1348–1355.https://doi.org/10.1038/nn.3470 - Looking on the (B)right side of life: cognitive ability and miscalibrated financial expectationsPersonality & Social Psychology Bulletin 51:1703–1715.https://doi.org/10.1177/01461672231209400 - IQ in high school as a predictor of midlife alcohol drinking patternsAlcohol and Alcoholism 59:agae035.https://doi.org/10.1093/alcalc/agae035 - BookChapter 10: alcohol dependenceIn: McManus S, Bebbington P, Jenkins R, Brugha T, editors. Mental Health and Wellbeing in England: Adult Psychiatric Morbidity Survey 2014. NHS Digital. pp. 10–17. - A distributed brain network predicts general intelligence from resting-state human neuroimaging dataPhilosophical Transactions of the Royal Society B 373:20170284.https://doi.org/10.1098/rstb.2017.0284 - BookHuman connectome projectIn: Jaeger D, Jung R, editors. Encyclopedia of Computational Neuroscience. Springer. pp. 1647–1650.https://doi.org/10.1007/978-1-0716-1006-0_592 - A comparative study of fault density prediction in aspect-oriented systems using MLP, RBF, KNN, RT, DENFIS and SVR modelsArtificial Intelligence Review 42:695–703.https://doi.org/10.1007/s10462-012-9348-9 - Cognitive functioning in prodromal psychosis: a meta-analysisArchives of General Psychiatry 69:562–571.https://doi.org/10.1001/archgenpsychiatry.2011.1592 - Development, reliability and factor analysis of a self-administered questionnaire which originates from the World Health Organization’s Composite International Diagnostic Interview - Short Form (CIDI-SF) for assessing mental disordersClinical Practice and Epidemiology in Mental Health 4:8.https://doi.org/10.1186/1745-0179-4-8 - Anxiety and cognitive functioning in the Maastricht study: a cross-sectional population studyJournal of Affective Disorders 319:570–579.https://doi.org/10.1016/j.jad.2022.09.072 - Memory and cognition in schizophreniaMolecular Psychiatry 24:633–642.https://doi.org/10.1038/s41380-018-0231-1 - Wechsler Adult Intelligence Scale IV (WAIS IV): return of the gold standardApplied Neuropsychology 16:85–87.https://doi.org/10.1080/09084280802644466 - Emotion and cognition interactions in PTSD: a review of neurocognitive and neuroimaging studiesFrontiers in Integrative Neuroscience 6:89.https://doi.org/10.3389/fnint.2012.00089 - Ridge regression: a historical contextTechnometrics 62:420–425.https://doi.org/10.1080/00401706.2020.1742207 - Research domain criteria (RDoC): toward a new classification framework for research on mental disordersThe American Journal of Psychiatry 167:748–751.https://doi.org/10.1176/appi.ajp.2010.09091379 - The relation between mood, cognition and psychosocial functioning in psychiatric disordersEuropean Neuropsychopharmacology 22 Suppl 3:S499–S504.https://doi.org/10.1016/j.euroneuro.2012.08.002 - The effect of splitting on random forestsMachine Learning 99:75–118.https://doi.org/10.1007/s10994-014-5451-2 - Model-based analysis of multishell diffusion MR data for tractography: how to get over fitting problemsMagnetic Resonance in Medicine 68:1846–1855.https://doi.org/10.1002/mrm.24204 - A global optimisation method for robust affine registration of brain imagesMedical Image Analysis 5:143–156.https://doi.org/10.1016/s1361-8415(01)00036-6 - The g factor: psychometrics and biologyNovartis Foundation Symposium 233:37–47.https://doi.org/10.1002/0470870850.ch3 - Consistent impaired verbal memory in PTSD: a meta-analysisJournal of Affective Disorders 111:74–82.https://doi.org/10.1016/j.jad.2008.02.007 - Why is cognitive ability associated with psychological distress and wellbeing? Exploring psychological, biological, and social mechanismsPersonality and Individual Differences 192:111592.https://doi.org/10.1016/j.paid.2022.111592 - Intelligence and substance useReview of General Psychology 14:382–396.https://doi.org/10.1037/a0021526 - The World Health Organization Composite International Diagnostic Interview short‐form (CIDI‐SF)International Journal of Methods in Psychiatric Research 7:171–185.https://doi.org/10.1002/mpr.47 - Mind matters: exploring the intersection of psychological factors and cognitive abilities of university students by using ANN modelNeuropsychiatric Disease and Treatment 20:137–148.https://doi.org/10.2147/NDT.S436975 - Age differences in short-term retention of rapidly changing informationJournal of Experimental Psychology 55:352–358.https://doi.org/10.1037/h0043688 - Neurocognitive deficits in depression: a systematic review of cognitive impairment in the acute and remitted stateEuropean Archives of Psychiatry and Clinical Neuroscience 273:1105–1128.https://doi.org/10.1007/s00406-022-01479-5 - The patient health questionnaire somatic, anxiety, and depressive symptom scales: a systematic reviewGeneral Hospital Psychiatry 32:345–359.https://doi.org/10.1016/j.genhosppsych.2010.03.006 - Social engagement and cognitive function in old ageExperimental Aging Research 35:45–60.https://doi.org/10.1080/03610730802545028 - An abbreviated PTSD checklist for use as a screening instrument in primary careBehaviour Research and Therapy 43:585–594.https://doi.org/10.1016/j.brat.2004.04.005 - A well-conditioned estimator for large-dimensional covariance matricesJournal of Multivariate Analysis 88:365–411.https://doi.org/10.1016/S0047-259X(03)00096-4 - Resting-state fMRI: a review of methods and clinical applicationsAJNR. American Journal of Neuroradiology 34:1866–1872.https://doi.org/10.3174/ajnr.A3263 - Is cognitive performance associated with anxiety and depression in first-episode psychosis?Journal of Affective Disorders 263:221–227.https://doi.org/10.1016/j.jad.2019.11.161 - FS_SFS: A novel feature selection method for support vector machinesPattern Recognition 39:1333–1345.https://doi.org/10.1016/j.patcog.2005.10.006 - BookSupport vector machines and support vector regressionIn: López OAM, editors. Multivariate Statistical Machine Learning Methods for Genomic Prediction. Springer. pp. 337–378.https://doi.org/10.1007/978-3-030-89010-0_9 - Educational differences in alcohol consumption and heavy drinking: an age-period-cohort perspectiveDrug and Alcohol Dependence 186:36–43.https://doi.org/10.1016/j.drugalcdep.2017.12.046 - The effects of lifetime trauma exposure on cognitive functioning in midlifeJournal of Traumatic Stress 33:773–782.https://doi.org/10.1002/jts.22522 - Differences in risk factors for self-harm with and without suicidal intent: findings from the ALSPAC cohortJournal of Affective Disorders 168:407–414.https://doi.org/10.1016/j.jad.2014.07.009 - Cognitive impairment in schizophrenia: aetiology, pathophysiology, and treatmentMolecular Psychiatry 28:1902–1918.https://doi.org/10.1038/s41380-023-01949-9 - Neurocognition in first-episode schizophrenia: a meta-analytic reviewNeuropsychology 23:315–336.https://doi.org/10.1037/a0014708 - Cognitive abnormalities in posttraumatic stress disorderCurrent Opinion in Psychiatry 22:19–24.https://doi.org/10.1097/YCO.0b013e328314e3bb - MRI atlas of human white matterAJNR American Journal of Neuroradiology 27:1384–1385. - Research domain criteria: cognitive systems, neural circuits, and dimensions of behaviorDialogues in Clinical Neuroscience 14:29–37.https://doi.org/10.31887/DCNS.2012.14.1/smorris - The revival of the Gini importance?Bioinformatics 34:3711–3718.https://doi.org/10.1093/bioinformatics/bty373 - BookTransport on riemannian manifold for functional connectivity-based classificationIn: Golland P, Hata N, Barillot C, Hornegger J, Howe R, editors. Medical Image Computing and Computer-Assisted Intervention – MICCAI 2014. Springer International Publishing. pp. 405–412.https://doi.org/10.1007/978-3-319-10470-6_51 - The continuum of psychotic symptoms in the general population: a cross-national studySchizophrenia Bulletin 38:475–485.https://doi.org/10.1093/schbul/sbq099 - An introduction to diffusion tensor image analysisNeurosurgery Clinics of North America 22:185–196.https://doi.org/10.1016/j.nec.2010.12.004 - The adaptive brain: aging and neurocognitive scaffoldingAnnual Review of Psychology 60:173–196.https://doi.org/10.1146/annurev.psych.59.103006.093656 - Analysis of sub-anatomic diffusion tensor imaging indices in white matter regions of Alzheimer with MMSE scoreComputer Methods and Programs in Biomedicine 117:13–19.https://doi.org/10.1016/j.cmpb.2014.06.004 - Using the Standardized Root Mean Squared Residual (SRMR) to assess exact fit in structural equation modelsEducational and Psychological Measurement 81:110–130.https://doi.org/10.1177/0013164420926231 - BookBayesian model selection and averagingIn: Friston KJ, Ashburner JT, Kiebel SJ, Nichols TE, Penny WD, editors. Statistical Parametric Mapping: The Analysis of Functional Brain Images. Academic Press, Ltd. pp. 454–467.https://doi.org/10.1016/B978-012372560-8/50035-8 - Childhood trauma is associated with poorer cognitive performance in older adultsThe Journal of Clinical Psychiatry 79:16m11021.https://doi.org/10.4088/JCP.16m11021 - The dunedin multidisciplinary health and development study: overview of the first 40 years, with an eye to the futureSocial Psychiatry and Psychiatric Epidemiology 50:679–693.https://doi.org/10.1007/s00127-015-1048-8 - Potential of radial basis function based support vector regression for global solar radiation predictionRenewable and Sustainable Energy Reviews 39:1005–1011.https://doi.org/10.1016/j.rser.2014.07.108 - Integrating across neuroimaging modalities boosts prediction accuracy of cognitive abilityPLOS Computational Biology 17:e1008347.https://doi.org/10.1371/journal.pcbi.1008347 - The alcohol use disorders identification test: an update of research findingsAlcoholism, Clinical and Experimental Research 31:185–199.https://doi.org/10.1111/j.1530-0277.2006.00295.x - How does it STAC up? Revisiting the scaffolding theory of aging and cognitionNeuropsychology Review 24:355–370.https://doi.org/10.1007/s11065-014-9270-9 - Cognitive deficits in patients with depressionThe Journal for Nurse Practitioners 14:437–443.https://doi.org/10.1016/j.nurpra.2018.03.006 - Evolution of support vector machine and regression modeling in chemoinformatics and drug discoveryJournal of Computer-Aided Molecular Design 36:355–362.https://doi.org/10.1007/s10822-022-00442-9 - Intelligence, alcohol consumption, and adverse consequences. A study of young Norwegian menScandinavian Journal of Public Health 49:411–418.https://doi.org/10.1177/1403494820944719 - Lavaan: an r package for structural equation modelingJournal of Statistical Software 48:1–36.https://doi.org/10.18637/jss.v048.i02 - Genome-wide association study meta-analysis of the Alcohol Use Disorders Identification Test (AUDIT) in two population-based cohortsThe American Journal of Psychiatry 176:107–118.https://doi.org/10.1176/appi.ajp.2018.18040369 - Evaluating the fit of structural equation models: tests of significance and descriptive goodness-of-fit measuresMethods of Psychological Research 8:23–74. - The random forest algorithm for statistical learningThe Stata Journal 20:3–29.https://doi.org/10.1177/1536867X20909688 - Estimating the dimension of a modelThe Annals of Statistics 6:461–464.https://doi.org/10.1214/aos/1176344136 - Time-varying functional connectivity predicts fluctuations in sustained attention in a serial tapping taskCognitive, Affective, & Behavioral Neuroscience 24:111–125.https://doi.org/10.3758/s13415-024-01156-1 - Commonality analysis: a method for decomposing explained variance in multiple regression analysesHuman Communication Research 5:355–365.https://doi.org/10.1111/j.1468-2958.1979.tb00649.x - Cognitive deficits in psychotic disorders: a lifespan perspectiveNeuropsychology Review 28:509–533.https://doi.org/10.1007/s11065-018-9388-2 - The default mode network in cognition: a topographical perspectiveNature Reviews. Neuroscience 22:503–513.https://doi.org/10.1038/s41583-021-00474-4 - Fast robust automated brain extractionHuman Brain Mapping 17:143–155.https://doi.org/10.1002/hbm.10062 - A tutorial on support vector regressionStatistics and Computing 14:199–222.https://doi.org/10.1023/B:STCO.0000035301.49549.88 - A hitchhiker’s guide to diffusion tensor imagingFrontiers in Neuroscience 7:31.https://doi.org/10.3389/fnins.2013.00031 - Bayesian measures of model complexity and fitJournal of the Royal Statistical Society Series B 64:583–639.https://doi.org/10.1111/1467-9868.00353 - Prediction of neurocognition in youth from resting state fMRIMolecular Psychiatry 25:3413–3421.https://doi.org/10.1038/s41380-019-0481-6 - BookChapter 2: common mental disordersIn: McManus S, Bebbington P, Jenkins R, Brugha T, editors. Mental Health and Wellbeing in England: Adult Psychiatric. Morbidity Survey 2014. NHS Digital. pp. 37–62. - Random forest: a classification and regression tool for compound classification and QSAR modelingJournal of Chemical Information and Computer Sciences 43:1947–1958.https://doi.org/10.1021/ci034160g - A review of structural magnetic resonance neuroimagingJournal of Neurology, Neurosurgery & Psychiatry 75:1235–1244.https://doi.org/10.1136/jnnp.2003.032714 - Current clinical applications of diffusion-tensor imaging in neurological disordersJournal of Clinical Neurology 14:129–140.https://doi.org/10.3988/jcn.2018.14.2.129 - The organization of the human cerebral cortex estimated by intrinsic functional connectivityJournal of Neurophysiology 106:1125–1165.https://doi.org/10.1152/jn.00338.2011 - Regression shrinkage and selection via the lassoJournal of the Royal Statistical Society Series B 58:267–288.https://doi.org/10.1111/j.2517-6161.1996.tb02080.x - Shrinkage priors for Bayesian penalized regressionJournal of Mathematical Psychology 89:31–50.https://doi.org/10.1016/j.jmp.2018.12.004 - BookThe support vector method of function estimationIn: Suykens JAK, Vandewalle J, editors. Nonlinear Modeling: Advanced Black-Box Techniques. Springer. pp. 55–85.https://doi.org/10.1007/978-1-4615-5703-6_3 - Multiple regression in behavioral research: explanation and predictionPersonnel Psychology 51:223–226. - Dorsal and ventral attention systemsThe Neuroscientist 20:150–159.https://doi.org/10.1177/1073858413494269 - Adult health status of women with histories of childhood abuse and neglectThe American Journal of Medicine 107:332–339.https://doi.org/10.1016/s0002-9343(99)00235-1 - Depression and cognitive impairment: current understanding of its neurobiology and diagnosisNeuropsychiatric Disease and Treatment 18:2783–2794.https://doi.org/10.2147/NDT.S383093 - A general cognitive ability factor for the UK BiobankBehavior Genetics 53:85–100.https://doi.org/10.1007/s10519-022-10127-6 - PLS-regression: a basic tool of chemometricsChemometrics and Intelligent Laboratory Systems 58:109–130.https://doi.org/10.1016/S0169-7439(01)00155-1 - Cognitive impairment in generalized anxiety disorder revealed by event-related potential N270Neuropsychiatric Disease and Treatment 11:1405–1411.https://doi.org/10.2147/NDT.S84666 - Mode of anisotropy reveals global diffusion alterations in attention-deficit/hyperactivity disorderJournal of the American Academy of Child and Adolescent Psychiatry 55:137–145.https://doi.org/10.1016/j.jaac.2015.11.011 - Segmentation of brain MR images through a hidden Markov random field model and the expectation-maximization algorithmIEEE Transactions on Medical Imaging 20:45–57.https://doi.org/10.1109/42.906424 - Educational attainment and drinking behaviors: mendelian randomization study in UK BiobankMolecular Psychiatry 26:4355–4366.https://doi.org/10.1038/s41380-019-0596-9 - The association between happiness and cognitive function in the UK BiobankCurrent Psychology 43:1816–1825.https://doi.org/10.1007/s12144-023-04446-y - Regularization and variable selection via the elastic netJournal of the Royal Statistical Society Series B 67:301–320.https://doi.org/10.1111/j.1467-9868.2005.00503.x Article and author information Author details Funding Health Research Council of New Zealand (21/618) - Jeremiah D Deng - Narun Pat Neurological Foundation of New Zealand (2350 PRG) - Jeremiah D Deng - Narun Pat Ministry of Business, Innovation and Employment (UOA2421) - Narun Pat University of Otago (The University of Otago Doctoral Scholarship) - Irina Buianova Health Research Council of New Zealand (24/838) - Jeremiah D Deng - Narun Pat Ministry of Business, Innovation and Employment (RTVU2403) - Narun Pat The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication. Acknowledgements We are thankful to Professor Bruce Russell for his assistance with access to the UK Biobank dataset. NP and JD were supported by Health Research Council of New Zealand (grant numbers 21/618 and 24/838) and by Neurological Foundation of New Zealand (grant number 2350 PRG). NP was supported by the Ministry of Business, Innovation and Employment (grant numbers UOA2421 and RTVU2403). IB was supported by the University of Otago. Ethics Version history - Preprint posted: - Sent for peer review: - Reviewed Preprint version 1: - Reviewed Preprint version 2: - Version of Record published: Cite all versions You can cite all versions using the DOI https://doi.org/10.7554/eLife.108109. This DOI represents all versions, and will always resolve to the latest one. Copyright © 2025, Buianova et al. This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited. Metrics - - 1,017 - views - - 60 - downloads - - 0 - citations Views, downloads and citations are aggregated across all versions of this paper published by eLife.