BACKGROUND:Treatment-resistant schizophrenia (TRS) affects 20-30% of individuals with schizophrenia, with persistent symptoms, functional impairment, and reduced quality of life. Clinical identification remains dependent on sequential antipsychotic trials despite reported structural brain differences between TRS and treatment-responsive schizophrenia (TxR). This study evaluated whether structural MRI features could discriminate clinically defined TRS from TxR using machine learning. METHODS:A total of 225 participants (122 TRS, 103 TxR) from multi-site studies in Canada and Japan were included. Cortical thickness, brain volume, surface area, and intrinsic curvature were derived from T1-weighted MRI using FreeSurfer, with feature engineering generating volumetric-cortical thickness interaction terms. A voting ensemble was evaluated under a primary leakage-controlled NeuroComBat harmonization and a secondary exploratory full-dataset harmonization. Performance was assessed using ROC-AUC, F1 score, precision, and recall. RESULTS:In the leakage-controlled analysis, the voting ensemble achieved a held-out ROC-AUC of 0.57 and macro F1 of 0.56; training cross-validation yielded ROC-AUC of 0.715. The full-dataset harmonization analysis yielded ROC-AUC of 0.60 and macro F1 of 0.62, interpreted cautiously due to leakage. Temporal-occipital cortical thickness and choroid plexus volume interaction terms contributed most to model performance. CONCLUSION:Structural MRI features may support cross-sectional discrimination of TRS from TxR. The divergence between harmonization strategies highlights the importance of leakage-aware preprocessing in neuroimaging. Validation in independent cohorts is required to establish whether these features contribute to earlier identification of treatment resistance.
Suicide represents a global public health crisis, claiming over 700,000 lives worldwide every year. Deep understanding of the neurobiological mechanisms will enable objective risk assessment and more effective prevention of suicide. Recent advances in neuroimaging have revealed that suicidal thoughts and behaviors are associated with disruptions in both structural and functional brain connectome organization. In addition, brain connectome profiles may represent "fingerprints" that capture individual variability in suicidality, offering a transformative framework for personalized assessment. In this review, we discuss the latest developments in connectome-based modeling of the suicidal brain, which can not only inform neurobiological mechanisms but also help to advance clinical translation. We first summarize previous MRI-based connectomic findings, emphasizing convergent evidence for disrupted prefrontal-limbic-subcortical circuitry and reduced global network integration as core features of suicidality. Next, we review PET and EEG/MEG studies to highlight future directions for multimodal and integrative connectomic research. With a specific focus on individualized application, we also highlight connectome-based machine learning findings and propose a novel paradigm using normative models to develop a connectome-based suicide risk calculator. Finally, we present current challenges and future directions to improve brain connectome research in suicidality, emphasizing the imperative of using high-quality longitudinal cohorts for validation.
BACKGROUND:Cortical gray matter loss is a common finding in magnetic resonance imaging (MRI) studies of psychosis and can progress with ongoing illness. A major unresolved question concerns whether these changes are driven by the illness or represent iatrogenic effects of antipsychotics. METHODS:In a triple-blind, randomized, placebo-controlled MRI study, 62 antipsychotic-naïve people with first-episode psychosis (FEP) received a second-generation antipsychotic or placebo over 6 months (n = 35 at 12 months) alongside a healthy control group (n = 27 at baseline, n = 21 at 12 months). T1-weighted scans were collected at baseline, 3 months, and 12 months. Linear mixed-effects models fitted to >160,000 cortical loci examined illness- and antipsychotic-related thickness changes. We also examined whether cortical changes were enriched within functional networks or cytoarchitectonic classes and whether they spatially correlated with normative positron emission tomography receptor/transporter densities and transcriptomically imputed cell densities. RESULTS:Over 12 months, the placebo group showed widespread cortical thinning compared with the control group (false discovery rate [FDR]-corrected p < .05), with the largest effects in frontal, cingulate, and occipital areas. No significant difference was detected between patients treated with antipsychotics and control participants. Thinning was not concentrated within specific functional networks, but highly differentiated koniocortical areas were relatively protected (pspin < .05 FDR-corrected). Thinning spatially aligned with normative distributions of GABAA/BZ, 5-HT1B, 5-HT2A, and H3 receptors (0.17 <r < 0.28; pspin < .05 FDR-corrected). No associations between thinning and symptom change were identified. CONCLUSIONS:Widespread cortical thinning occurs over the first year of FEP in people not receiving antipsychotics. No such thinning is evident in patients receiving antipsychotics. These findings suggest that cortical thinning in early psychosis is an illness-related phenomenon.
BACKGROUND:Response to repetitive transcranial magnetic stimulation (rTMS) in major depressive disorder (MDD) varies substantially. Normative modeling of functional connectivity can disentangle disease-related pathophysiology from demographic variability, potentially refining personalized targeting. METHODS:We constructed a normative model of subgenual anterior cingulate cortex (sgACC)-dorsolateral prefrontal cortex (DLPFC) functional connectivity using exclusively healthy controls (HCs; DIRECT dataset, n = 1,313). Individual Z-score maps were generated for 1,583 MDD patients; the DLPFC voxel with the most negative Z-score defined the functional connectivity normative deviation (FCND)-guided target. Clinical utility was tested in two independent rTMS cohorts (active: n=39, sham: n=23; accelerated iTBS: n=15) by correlating Euclidean distance from the stimulation site to the FCND target with improvement. We explored whether Z-score normalization statistically mediated this relationship. RESULTS:In the DIRECT dataset, patients with MDD showed significantly more negative values for the most negative Z-score (Min Z-score) than HCs, and this Min Z-score correlated negatively with depressive symptom severity, as measured by the 17-item Hamilton Depression Rating Scale. In the YT active group, shorter distance to the FCND-guided target was associated with greater clinical improvement, while no association was observed in the sham group. This association was also significant in the aiTBS dataset. Exploratory mediation analysis revealed an indirect effect statistically consistent with the hypothesis that distance influences improvement through Z-score normalization in the active group, with a directionally consistent trend in the iTBS dataset. CONCLUSION:The FCND-guided target framework provides a personalized, mechanism-informed strategy for precision neuromodulation in MDD.
Autism presents significant challenges for clinical research and biomarker development due to behavioral heterogeneity and reliance on clinician- and caregiver-report measures that are limited in sensitivity, objectivity, and scalability. Biomarkers offer a complementary approach by providing biologically grounded, standardized, and potentially more sensitive indicators of the autism phenotype. We outline the potential advantages of biomarkers over traditional clinical assessments, operationalize biomarkers using FDA/NIH consensus frameworks, and evaluate recent progress from large-scale biomarker consortia and remaining challenges for two well-developed measurement modalities: electroencephalography (EEG) and eye-tracking (ET). We summarize findings from one of these recent biomarker consortia, the Autism Biomarkers Consortium for Clinical Trials (ABC-CT), that has established the feasibility, stability, and regulatory relevance of several candidate biomarkers. Despite meaningful advances in reproducibility of biological factors associated with autism, clinically actionable biomarkers remain elusive. We describe potential paths forward to overcome this obstacle, including leveraging neurogenetic syndromes and genomic stratification, aligning biomarkers with biologically specified behavioral constructs, expanding measurement into motor and vestibular domains, and applying data-driven, multimodal, and digital phenotyping approaches to advance clinical trial readiness in autism.
BACKGROUND:Episodic memory processes and their underlying neural circuitry have been closely linked with stress-related psychopathology; however, the temporal and directional nature of these links remains unclear. Understanding how large-scale brain networks and psychopathology symptoms relate to changes in episodic memory across development is important for identifying at-risk youth and guiding developmentally-tailored interventions. METHODS:We used latent change score modeling to examine longitudinal associations between stress-related psychopathology and episodic memory in N = 3,466 youth aged 9-10 from the Adolescent Brain Cognitive Development (ABCD) Study. We then trained and tested resting-state functional connectivity (rsFC) networks to predict changes in episodic memory using connectome-based predictive modeling, and examined how these networks and stress-related psychopathology mediate links between early-life stress and memory. RESULTS:Stress-related psychopathology emerged as a leading predictor of change in episodic memory in youth. Whole-brain rsFC networks successfully predicted developmental changes in episodic memory. Specifically, stronger connectivity in the visual association area, primary visual cortex, and secondary visual cortex predicted less improvement in episodic memory over two years, whereas stronger connectivity in the cerebellum and subcortical networks predicted greater improvement in episodic memory over two years. Finally, rsFC and stress-related psychopathology mediated the association between higher early-life stress and less improvement in episodic memory. CONCLUSIONS:Together, these findings suggest that early-life stress, network connectivity, and stress-related psychopathology contribute to episodic memory development. The identification of these neurobiological and early environmental predictors may contribute to better understanding of potential interventions targeting maladaptive memory processes in youth with stress-related psychopathology.
Precision mental health aims to enable personalized care for mental disorders via the identification of brain-behavior associations at the individual level, yet current frameworks often face challenges in generalizability, interpretability, and clinical translation. Contrastive machine learning (CML) has emerged as a promising paradigm for characterizing individual brain variations by extracting brain dimensions that capture both disease-relevant aberrations and inter-subject heterogeneity. In this Review, we synthesize the conceptual foundations and recent methodological advances of CML. We compare CML with related frameworks and illustrate how it integrates with subtyping and predictive modeling pipelines, situating it within the broader landscape of precision mental health. We then review emerging applications that use CML to link brain structure and function to cognition, emotion, and treatment response. Finally, we outline future directions of applications and methodological innovations where CML could further advance personalized diagnosis and intervention in mental health.
