Motor performance (MP) is essential for maintaining functional independence, particularly in later life. However, the relationship between MP and sleep quality, depressive symptoms, and their underlying brain substrates remains obscure. We employed four samples of younger/mid-to-older adults (n = 1,954) from the Human Connectome Project-Young Adult (HCP-YA), HCP-Aging (HCP-A), and enhanced Nathan Kline Institute-Rockland sample (eNKI-RS) to assess the replicability of our findings. Using canonical correlation analyses within a machine learning framework, we investigated the associations of sleep quality, depressive symptoms, and grey matter volume (GMV) with MP. In the combined model of the HCP-YA sample, a canonical variate of better sleep, mild, sub-clinical depressive symptoms, and altered GMV of several cortical (including precentral and fusiform gyrus), thalamus, and cerebellar brain regions was associated with a canonical variate of better MP (r = 0.2, SD = 0.05). This pattern was conceptually replicated in the young eNKI-RS sample (r = 0.25, SD = 0.13). In the HCP-A sample, a variate of better sleep quality, fewer depressive symptoms, and increased GMV was associated with a variate of MP (r = 0.18, SD = 0.1), but these findings did not replicate in the mid-to-older eNKI-RS sample (r = 0, SD = 0.12). Across all samples, variates of increased GMV were associated with variates of better MP, suggesting potential neuroanatomical underpinnings. We observed age-related variations in the multivariate associations between sleep quality, depressive symptoms, and GMV with MP.

BACKGROUND AND PURPOSE:Multi-echo (ME) functional MRI (fMRI) acquisition improves separation of signal from noise relative to single-echo (SE). We tested whether this enhances reliability of functional connectivity (FC), with a focus on personalizing transcranial magnetic stimulation (TMS) targets in the dorsolateral prefrontal cortex (DLPFC) in patients with depression. MATERIALS AND METHODS:Resting-state fMRI scans were acquired from adult patients with major depression (20 female, 15 male) presenting for clinical TMS using either SE (n=21) or ME (n=31). Each subject's fMRI timeseries was split in half, and voxel-wise seed-based FC was computed for 100 general regions of interest (ROIs) and for two TMS-specific ROIs: subgenual cingulate cortex (SGC) and a previously published depression circuit (DEP). Reliability was assessed using (1) spatial correlation between split-half connectivity maps and (2) intraclass correlation coefficient (ICC) for each ROI's connectivity to the DLPFC. RESULTS:In general ROI analysis, ME showed significantly higher whole-brain split-half correlations than SE (p = 0.006) and higher ICC (ΔICC = 0.16; p = 0.03). In TMS-specific ROI analysis, ME showed higher split-half correlations for both the SGC-DLPFC (p = 0.04) and DEP-DLPFC (p = 0.01). TMS-specific ICC values were numerically higher for ME (SGC-DLPFC: 0.47; DEP-DLPFC: 0.75) than for SE (0.02 and 0.40, respectively), although these differences were not statistically significant. CONCLUSION:ME fMRI improves general FC reliability over SE, with suggested advantages for TMS-specific measures. Future work is needed to determine whether these gains meaningfully improve TMS targeting.

Extensive neuroimaging research in temporal lobe epilepsy with hippocampal sclerosis (TLE-HS) has identified brain atrophy as a disease phenotype. While it is also related to a complex genetic architecture, the transition from genetic risk factors to brain vulnerabilities remains unclear. Using a population-based approach, we examined the associations between epilepsy-related polygenic risk for HS (PRS-HS) and brain structure in healthy developing children, assessed their relation to brain network architecture, and evaluated its correspondence with case-control findings in TLE-HS diagnosed patients relative to healthy individuals. We used genome-wide genotyping and structural T1-weighted MRI of 3826 neurotypical children from the Adolescent Brain Cognitive Development (ABCD) study. Surface-based linear models related PRS-HS to cortical thickness measures, and subsequently contextualized findings with structural and functional network architecture based on epicentre mapping approaches. Imaging-genetic associations were then correlated to atrophy and disease epicentres in 785 patients with TLE-HS relative to 1512 healthy controls aggregated across multiple sites. Higher PRS-HS was associated with decreases in cortical thickness across temporo-parietal as well as fronto-central regions of neurotypical children. These imaging-genetic effects were anchored to the connectivity profiles of distinct functional and structural epicentres. Compared with disease-related alterations from a separate epilepsy cohort, regional and network correlates of PRS-HS strongly mirrored cortical atrophy and disease epicentres observed in patients with TLE-HS and were highly replicable across different studies. Findings were consistent when using statistical models controlling for spatial autocorrelations and robust to variations in analytic methods. Capitalizing on recent imaging-genetic initiatives, our study provides novel insights into the genetic underpinnings of structural alterations in TLE-HS, revealing common morphological and network pathways between genetic vulnerability and disease mechanisms. These signatures offer a foundation for early risk stratification and personalized interventions targeting genetic profiles in epilepsy.

Functional connectivity (FC) is often used to identify personalized targets for transcranial magnetic stimulation (TMS). However, existing methods often overlook individual differences in whole-cortex network organization. Furthermore, in some personalized TMS protocols, lower stimulation intensity is used for targets closer to the scalp, which may improve patient tolerance. Here, we develop an algorithm to simultaneously optimize FC and scalp proximity for target localization. We first use the multi-session hierarchical Bayesian model (MS-HBM) to estimate high-quality individual-specific cortical networks. A tree-based algorithm is then used to select the optimal target. With essentially no parameter to tune, our framework may potentially improve generalizability across populations. We compare our approach with existing "cluster" and "cone" algorithms. In two test-retest datasets of healthy individuals from the United States and Singapore, tree-based MS-HBM reliably identifies personalized TMS targets for depression near the scalp. Tree-based MS-HBM targets compare favorably with cluster and cone targets in terms of reliability, scalp proximity, and FC to the subgenual anterior cingulate cortex (sACC) in new out-of-sample MRI sessions. To demonstrate versatility, the same algorithm identifies personalized anxiety targets without tuning any parameter. In patients with treatment-resistant depression, tree-based MS-HBM targets compare favorably with cluster and cone targets in terms of reliability, scalp proximity, and sACC FC, hypothetically reducing stimulation intensity by 15% and 5%, respectively. MS-HBM also exhibits the best (most negative) electric-field hotspot sACC FC and highest reliability in induced electric fields. Overall, tree-based MS-HBM provides a robust, generalizable framework to estimate near-scalp personalized targets across populations.

Brain atrophy may precede cognitive and functional impairment in Alzheimer's disease (AD), but at this "preclinical" stage, it remains unclear whether atrophy localizes to specific brain networks and whether such localization is associated with clinical outcomes. We investigated cortical thickness in 1778 cognitively unimpaired (CU) older adults with amyloid-β (Aβ) PET from the A4 (Anti-Amyloid Treatment in Asymptomatic Alzheimer's Disease) and LEARN (Longitudinal Evaluation of Amyloid Risk and Neurodegeneration) studies, with a subset (N = 445) with tau PET. We estimated the networks disrupted by each individual's cortical thinning using a large normative connectome database (N = 1000), and tested whether this preclinical AD atrophy network is associated with clinical manifestations and longitudinal cognitive (PACC) and functional (CDR) trajectories. Distinct networks connected to atrophy patterns were associated with Aβ and regional tau in CU older adults. These networks were similar to a previously published atrophy network for AD dementia (Aβ: r = 0.817, P = 0.006; tau: r = 0.712, P = 0.046). Atrophy connectivity to this preclinical AD network was associated with higher Aβ and tau, independent of total cortical atrophy, and cross-sectionally with lower cognition, greater subjective cognitive decline, and increased anxiety; longitudinally, it predicted faster cognitive and functional decline over ~5 years. After tau adjustment, the functional-decline effect was preserved while the cognitive-slope effect was largely attenuated. Atrophy in preclinical AD localizes to a network resembling AD dementia, and is independently associated with AD pathologies, clinical outcomes, and longitudinal decline. Network-level neurodegeneration is detectable and clinically informative in preclinical AD, supporting future network-based research and therapeutic development.

OBJECTIVE:This study was undertaken to reliably estimate individual-specific resting-state cortical networks and determine whether language network topography can predict task-based language dominance in drug-resistant epilepsy. METHODS:We utilized a multisession hierarchical Bayesian model (MS-HBM) trained on drug-resistant epilepsy patients to map high-quality individual-specific cortical networks in this population (n = 65) with only 6-24 min of resting-state functional magnetic resonance imaging (fMRI). We compared the quality of networks to MS-HBM models trained on healthy participants from the human connectome project (n = 40) and tested the generalizability of the model in an independent cohort of drug-resistant epilepsy participants (n = 26). Resting-state language network topography was then used to predict task-based language dominance. RESULTS:Ninety-one participants with drug-resistant epilepsy (National Institutes of Health, n = 65; University of Iowa, n = 26) were included: 61 (67.0%) temporal lobe epilepsy, 29 (31.9%) extratemporal lobe epilepsy, and one (1.1%) undetermined seizure onset zone. The mean age was 33.0 ± 11.4 years, and 50 (54.9%) were male. There were 40 healthy participants with a mean age of 29.0 ± 4.0 years, and 16 (40.0%) were male. MS-HBM trained on drug-resistant epilepsy estimated individual-specific networks that more accurately capture cortical functional organization than group-average networks or MS-HBM trained on healthy participants. The trained MS-HBM model generalized to an independent cohort of drug-resistant epilepsy participants with concurrent intracranial electrical stimulation and fMRI. Critically, cortical evoked fMRI activity aligned more closely with individual-specific networks than with group-average networks. Furthermore, individual-specific language network topography significantly predicted task-based language dominance, achieving high accuracy for left (area under the curve [AUC] = .82), bilateral (AUC = .72), and right (AUC = .83) dominance. SIGNIFICANCE:These results demonstrate that MS-HBM captures functionally meaningful network reorganization in drug-resistant epilepsy and enables accurate, individual-level prediction of language lateralization, with direct implications for presurgical functional mapping.

BACKGROUND:Executive function (EF) impairments are often seen in mental disorders, particularly schizophrenia (SZ), where they relate to adverse outcomes. As a heterogeneous construct, how specifically each dimension of EF to characterize the diagnostic and prognostic aspects of SZ remains opaque. STUDY DESIGN:We used classification models with a stacking approach on systematically measured EFs using 6 tasks to discriminate 195 patients with SZ from healthy individuals. Baseline EF measurements were moreover employed to predict symptomatically remitted or non-remitted prognostic subgroups. EF feature importance was determined at the group-level and the ensuing individual importance scores were associated with 4 symptom dimensions. STUDY RESULTS:The models highlighted the importance of inhibitory control (interference and response inhibitions) or working memory (WM) in accurately identifying individuals with SZ (area under the curve [AUC] = 0.87) or those in remission (AUC = 0.81). Patients who are correctly classified, in the association with the contribution of interference inhibition function to our diagnostic classifier, present more severe baseline negative symptoms compared to those who are more likely to be misclassified. Also, linked to the function of WM updating, patients who are successfully classified as remitted display milder cognitive symptoms at follow-up. Remitted patients do not differ significantly from non-remitted cases in baseline EF assessments or overall symptom severity. CONCLUSIONS:Our work indicates that impairments in specific EF dimensions in SZ are differentially linked to individual symptom-load and prognostic outcomes. Thus, assessments and models based on EF may be promising in the clinical evaluation of this disorder.

Individual differences in neural circuits underlying emotional regulation, motivation, and decision-making are implicated in many psychiatric illnesses. Interindividual variability in these circuits may manifest, at least in part, as individual differences in impulsivity at both normative and clinically significant levels. Impulsivity reflects a tendency towards rapid, unplanned reactions to internal or external stimuli without considering potential negative consequences coupled with difficulty inhibiting responses. Here, we use multivariate brain-based predictive models to explore the neural bases of impulsivity across multiple behavioral scales, neuroanatomical features (cortical thickness, surface area, and gray matter volume), and sexes (females and males) in a large sample of youth from the Adolescent Brain Cognitive Development (ABCD) Study at baseline (n = 9,099) and two-year follow-up (n = 6,432). Impulsivity is significantly associated with neuroanatomical variability, and these associations vary across behavioral scales and neuroanatomical features. Impulsivity broadly maps onto cortical thickness in dispersed regions (e.g., inferior frontal, lateral occipital, superior frontal, entorhinal), as well as surface area and gray matter volume in specific medial (e.g., parahippocampal, cingulate) and polar (e.g., frontal and temporal) territories. Importantly, while many relationships are stable across sexes, others are sex-specific. These results highlight the complexity of the relationships between neuroanatomy and impulsivity across scales, features, sexes, and time points in youth. These findings suggest that neuroanatomy, in combination with other biological and environmental factors, reflects a key driver of individual differences in impulsivity in youth. As such, neuroanatomical markers may help identify youth at increased risk for developing impulsivity-related illnesses. Furthermore, this work emphasizes the importance of adopting a multidimensional and sex-specific approach in neuroimaging and behavioral research.

Negative symptoms of schizophrenia (SCZ), particularly amotivation, are prominent across both SCZ and bipolar disorder (BD). While orbitofrontal cortex (OFC) alterations have been implicated in the development of negative symptoms, their contributions across disorders remain to be established. Here, we examined how OFC thickness and network associations relate to amotivation compared to diminished expression across the BD-SCZ spectrum. We included 50 individuals with SCZ, 49 with BD, and 122 controls. We assessed amotivation and diminished expression and estimated thickness in the medial and lateral OFC as regions of interest as well as 64 other cortical regions. Across BD and SCZ, reduced right lateral and bilateral medial OFC thickness were specifically associated with amotivation, but not diminished expression or other clinical factors. We then generated intra-individual OFC structural covariance networks to evaluate how the system-level embedding of the OFC would link to brain-wide cortical maps of negative symptoms. We found that medial OFC covariance networks spatially correlated with the brain-wide cortical alterations of both negative symptom dimensions. Further analyses in independent SCZ data from the ENIGMA consortium (n = 4474) revealed associations with lateral OFC covariance networks. Finally, the brain-wide cortical alterations of amotivation were significantly correlated with normative functional and structural white-matter connectivity profiles of the right medial and left lateral OFC as well as adjacent prefrontal and limbic regions. Our work identifies OFC alterations as a possible transdiagnostic signature of amotivation and provides insights into network associations underlying the system-wide cortical alterations of negative symptoms across SCZ and BD.

Censoring high-motion volumes in fMRI is common practice to reduce effects of head motion on functional connectivity (FC). Although aggressive censoring removes more noise, it causes extensive data loss, creating a tradeoff that may ultimately improve or degrade FC accuracy. Here, we evaluate how censoring affects FC estimation and downstream brain-wide association studies (BWAS). Using extensively sampled participants from the Human Connectome Project (HCP) Retest dataset, we establish individual "ground truth" FC and assess the accuracy of FC estimated from 5-30 minute scans. We find that censoring degrades FC accuracy, with more aggressive censoring being more detrimental, particularly among participants exhibiting above-average motion. In these participants, aggressive censoring reduces FC accuracy by 30

Machine learning is accelerating biomedical research. Cross-validation is widely used to compare predictive performance - not only to benchmark algorithms, but also to inform scientific applications, such as ranking biomarkers. However, prediction performance estimates across cross-validation folds are not independent. Standard tests for comparing prediction performance (e.g., paired t-test) assume independence and can therefore inflate false positive rates. In a PRISMA-guided meta-analysis of 210 studies (impact factor ≥15, 1 June 2020 - 1 June 2025), we find that 97% ignored fold dependence when comparing prediction performance. This problem is ubiquitous across scientific fields and unaffected by impact factor, rigor-promoting policies, or open science practices. Simulations across 420 scenarios spanning four diverse datasets show that ignoring fold dependence leads to invalid false positive control in most settings. Repeated cross-validation further compounds this problem, with false positive rates rising toward 100% as the number of repetitions grows. Existing fold-dependence-aware tests rely on strong assumptions because the variance of fold-level statistics and the between-fold correlation cannot be disentangled under standard cross-validation. We therefore propose the SHARP (Split-HAlf RePeated) test, a simple modification to standard cross-validation that enables direct estimation of variance and correlation. Benchmarked against 12 tests, SHARP provides the best overall balance of false-positive control, statistical power, and confidence-interval calibration across simulation schemes. We conclude by providing best practices and reporting guidelines for valid model comparison inference in biomedical machine learning and beyond.

Anhedonia is a core feature of major depressive disorder (MDD), yet links between peripheral molecular signatures and cortical network architecture remain poorly defined. We enrolled 210 participants, including 56 unmedicated MDD patients with high-anhedonia (HA), 61 with low-anhedonia (LA), and 93 healthy controls (HC). Morphometric similarity networks (MSNs) from structural MRI were compared between HA and LA. MSNs index individual-level network organization by quantifying inter-regional morphometric similarity. Regional MSN patterns were linked to Allen Human Brain Atlas using Spearman correlations with spin tests and a multi-K stability screen. Whole-blood RNA-seq (n = 199) was integrated with MSN features via sparse partial least squares-canonical correlation (sPLS-C), with key blood analyses repeated after leukocyte-composition adjustment. Gene Ontology over-representation and MAGMA gene-level analyses provided pathway context. HA showed greater MSN integration than LA, particularly within default-mode and somatomotor networks. MSN maps were negatively correlated with dopamine-transporter and kappa-opioid-receptor densities. Imaging-derived gene associations were enriched for regulation of Toll-like-receptor-3 signaling. In blood, sPLS-C revealed coupling between MSN features and a transcriptomic signature enriched for T-cell activation/differentiation and lymphocyte-apoptosis pathways. After composition adjustment, the pre-specified blood signature did not differ between HA and LA, indicating that between-group differences were largely composition-driven. As supportive genetic context, over-representation on MAGMA FDR-significant genes suggested protocadherin-mediated homophilic adhesion. Peripheral immune-redox pathway enrichment aligns with anhedonia-related cortical network alterations, whereas between-group blood differences are chiefly composition-driven. Adjusting for blood-cell composition is essential, this multimodal framework nominates immune-modulatory/redox targets and synaptic-adhesion biology for precision stratification and intervention.

BrainAge models hold promise as a clinical biomarker for developmental brain health, especially in childhood when there is the potential for early intervention. To distinguish between normative developmental variance and pathological divergence, BrainAge models should reflect the dynamic and diverse neurodevelopmental processes that occur in distinct developmental windows across childhood. We utilized multi-modal neuroimaging data from three pediatric cohorts covering ages 4 to 13 years (n = 1005, 2126 scans), split into Train and Test datasets. Twelve sex-stratified BrainAge models were built stratified by type and different combinations of neuroimaging features. Model types were "Full-Span" models covering the full age range, and "Phase-Specific" models split into early- and late-childhood. We first compared BrainAge estimates in the Test dataset amongst our candidate models, then benchmarked the best-performing model against published pre-trained models and DNA-based biological age measures. Our findings show that a BrainAge model that was phase-specific and consisted of both structural and functional features (cortical thickness, subcortical volumes, and functional network integration measures) showed good prediction of age and best distinguished between healthy and symptomatic subgroups. We present a proof-of-concept for developmental models supporting building BrainAge models of higher temporal resolution that align to different childhood developmental phases.

Abstract Background and Hypothesis Schizophrenia (SZ) and bipolar disorder (BD) share overlapping yet distinct clinical profiles and system-wide brain alterations. Macroscale functional connectivity gradients capture principal axes of cortical organization, including the separation of unimodal and transmodal systems, offering a low-dimensional lens on individual differences in brain architecture. Whether these axes reflect shared or diagnosis-specific variation across the SZ-BD spectrum is unknown. Study Design Using resting-state fMRI from 187 adults (110 HC, 37 SZ, 40 BD) from the UCLA Consortium for Neuropsychiatric Phenomics, we derived individual low-dimensional gradients and applied three analyses: case-control comparisons at both the cortical network and subcortical region-of-interest level, Partial Least Squares (PLS) regression linking gradients to clinical phenotypes, and individual-level similarity indices (SI-PLS) positioning participants within a gradient–behaviour space. Study Results While the gradient structure (G1: visual-somatomotor and G2: unimodal-transmodal) was preserved across groups, patient groups showed greater deviations along both axes. Network analyses revealed transdiagnostic frontoparietal compression in G2, alongside disorder-specific effects: visual pole contraction and subcortical amygdala displacement in SZ, and somatomotor displacement in BD. PLS identified a BD-associated profile of preserved gradient architecture and lower symptom burden, contrasting with an SZ-associated profile of greater cognitive impairment and symptom severity. SI-PLS scores placed SZ and BD in distinct regions of a shared two-dimensional neural space, with HC between them. Conclusions Differences across the SZ–BD spectrum organize along two principal axes, revealing transdiagnostic alterations in higher-order association systems alongside disorder-specific sensory signatures. These findings support a multi-axis dimensional framework for understanding clinical heterogeneity in psychosis.

Longitudinal dementia progression prediction is essential for clinical decision-making. However, models often degrade on external cohorts due to systemic missingness - where certain biomarkers available during training are completely absent at test time - compounded by distribution shifts and patient-specific variability. Here, we propose Progression-aware Feature Fusion with Test-Time Adaptation (ProFuse-TTA), a two-stage hierarchical Transformer for longitudinal dementia prediction. Stage 1 learns per-biomarker temporal representations from irregular observations without imputation. Stage 2 fuses them via cross-feature attention, with simulated modality dropout during training for robustness to systemic missingness. At inference, a lightweight test-time adaptation module performs per-individual calibration. We trained on ADNI and evaluated on three external cohorts comprising 2,316 participants and 13,205 timepoints, with controlled modality ablation experiments isolating the effect of systemic missingness. We compared against six baselines, four from a recent benchmark study and two new baselines including one built on a tabular foundation model. ProFuse-TTA achieved the best cross-dataset performance in 8 of 9 settings across clinical diagnosis, MMSE, and hippocampal volume prediction, and ranked first in 14 of 15 ablation scenarios. The model maintained superior performance across varying input lengths and prediction horizons up to 6 years. Pretrained ADNI models are available at XXX.

Head motion systematically biases functional connectivity (FC) estimates in resting-state functional MRI (rs-fMRI). A common mitigation strategy is to censor high-motion volumes and discard high-motion runs. However, overly stringent censoring risks discarding signal alongside noise, potentially degrading FC estimates. Here, we test the efficacy of various censoring strategies on individual-specific cortical parcellations and personalized transcranial magnetic stimulation (TMS) target selection. Using precision-fMRI datasets comprising 50 individuals, we define individualized "ground-truth" references from ≥1 hour of low-motion data per participant. We then simulate 10-min or 20-min rs-fMRI sessions with varying motion levels from the remaining data, yielding final samples of 22 and 19 participants, respectively. Higher motion produces parcellations and TMS targets that deviate further from the ground-truth references. However, at any given motion level, lenient censoring produces higher quality parcellations and personalized TMS targets than strict censoring. The improvement is comparable to doubling scan duration from 20 to 40 min under strict censoring. With personalized connectome-guided TMS, a common dilemma is whether to rescan patients with only high-motion runs. A mixed-motion session with one low-motion run and one high-motion run may often be considered usable after discarding the high-motion run and strict censoring. We find that lenient censoring of high-motion-only sessions yields TMS targets comparable to - or even better than - those derived from strictly censored mixed-motion sessions. Therefore, within the motion range and parcellation/TMS targeting frameworks evaluated here, patients may not need to be re-scanned solely because all runs exceed strict censoring criteria.

Background. Artificial intelligence (AI) and related digital technologies are rapidly entering mental health care, psychosomatic medicine, and psychotherapy. Yet the clinical value of these tools cannot be judged by technical performance alone. The field is characterized by diagnostic heterogeneity, reliance on language and self-report, complex biopsychosocial causation, and substantial variability in response to interventions. Clinically meaningful progress therefore requires integration of biological markers with patients’ lived experience, personal history and context, alongside person-centered precision decision-making and intervention. Summary. This review examines how AI may contribute to mental health care across multiple levels, including diagnostic assessment, biomarker development, treatment selection, AI-enabled non-invasive neuromodulation, closed-loop adaptation, and emerging chatbot-based interventions. We argue that AI should support and improve, not replace, clinical assessments and care by integrating complex data into person-centered care, while raising challenges of validation, safety, equity, commodification and governance. Key Messages. AI can strengthen and improve clinically effective, context-sensitive, and person-centered mental health care. AI may enhance diagnosis, treatment selection, and therapeutic options while preserving the human-human therapeutic alliance. The future of AI-enabled mental health depends on combining computational precision with human judgment, ethical oversight, and clinical relevance to improve patient outcomes and quality of life. Keywords: Artificial intelligence; Mental health; Personalized treatment; Psychotherapy; Psychosomatic medicine

Abstract How functional brain networks and cognition co-evolve during adolescent development remains poorly understood. Using baseline and Year 2 data from 2949 individuals in the Adolescent Brain Cognitive Development Study, we trained kernel ridge regression models to predict cognitive ability from resting-state functional connectivity. We find that baseline functional connectivity more strongly predicts future cognitive ability than baseline cognitive ability. Models trained on baseline functional connectivity to predict baseline cognition generalize better to Year 2 functional connectivity and cognition, suggesting that brain–cognition relationships strengthen over time. Intriguingly, baseline functional connectivity outperforms longitudinal functional connectivity change in predicting future cognitive ability. While longitudinal functional connectivity change is less reliable than baseline functional connectivity – intraclass correlation coefficient 0.24 vs. 0.56 – shortening scan duration to reduce reliability of baseline functional connectivity does not eliminate the predictive gap. Furthermore, neither baseline functional connectivity nor functional connectivity change meaningfully predicts longitudinal change in cognitive ability. We also identify converging and diverging predictive network features across cross-sectional and longitudinal brain-cognition models – a multivariate twist on Simpson’s paradox – with clear sex-specific patterns. Overall, in early adolescence, stable individual differences in brain functional network organization play a more critical role than dynamic changes in shaping future cognitive outcomes.

Personalized connectivity-guided accelerated intermittent theta burst stimulation (iTBS), like the Stanford Accelerated Intelligent Neuromodulation Therapy (SNT), shows high efficacy for treatment-resistant depression (TRD) in Western cohorts. However, generalizability to other demographics with substantial comorbidity remains unclear. Here, we evaluate connectivity-guided iTBS in a naturalistic Asian TRD population with high comorbidity burden. Twenty TRD participants received 50 sessions of TAO-TMS (Tree-based Algorithm for Optimized Transcranial Magnetic Stimulation) over 5 days. Participants averaged 1.6 psychiatric comorbidities, including personality disorders, autism and obsessive-compulsive disorder. TAO-TMS personalizes targets within attentional networks and maximize anti-correlation with the subgenual anterior cingulate cortex. Its near-scalp targets reduce stimulation intensity under the SNT protocol, improving patient comfort. Clinical response was defined as ≥50% reduction in the Montgomery-Åsberg Depression Rating Scale within four weeks of treatment. TAO-TMS yielded 70% response rate. Among patients who met typical randomized-trial eligibility criteria (N = 11), response rate was 83%. For context, non-accelerated BeamF3 TMS at the same hospital historically achieved response rate of 21%, indicating a patient population profile less responsive to TMS than those recruited in typical clinical trials. Post-hoc electric-field modeling showed that TAO-TMS improved network focality by 21% over BeamF3 targets. Functional connectivity changes were significant within every participant, but highly heterogeneous across participants. TAO-TMS was more cost-effective than electroconvulsive therapy (ECT), saving US$37,838 with higher quality-adjusted life years (QALYs 0.69 vs 0.65). These findings provide early evidence for the generalizability of connectivity-guided personalized TMS in a naturalistic Asian TRD population with substantial psychiatric comorbidities. TAO-TMS offers a cost-effective alternative to ECT, positioning it as a viable precision psychiatry intervention.