Home | Publications | VPE+26

Neurobiological Profiles of Vulnerability and Functional Outcomes in Early-Stage Psychiatric Disorders: A Multi-Level Machine Learning Analysis

MCML Authors

Clara Sophie Vetter

→ Group Daniel Rückert
Artificial Intelligence in Healthcare and Medicine

Daniel Rückert

Prof. Dr.

Director

Artificial Intelligence in Healthcare and Medicine

Abstract

Introduction: Early-stage psychiatric disorders are highly heterogeneous and often associated with long-term functional disability. Despite advances in clinical staging, current risk stratification models insufficiently reflect the complex biological and psychosocial processes driving individual vulnerability. There is a critical need for data-driven approaches that integrate information across domains to identify transdiagnostic profiles of risk and resilience that are generalisable, interpretable, and clinically meaningful. Our study aimed to identify multimodal latent vulnerability signatures in early-stage psychiatric disorders and to evaluate their clinical relevance for stratifying risk of poor functional outcomes over time.<br>Methods: This study leveraged data from the longitudinal PRONIA cohort (N = 1,059), a multisite European study recruiting participants across ten centres in Finland, Germany, Italy, Switzerland, and the UK. Participants included help-seeking individuals meeting criteria for clinical high risk for psychosis (CHR), recent-onset depression (ROD), or recent-onset psychosis (ROP), as well as healthy controls. Multimodal baseline assessments included structural neuroimaging, polygenic risk scores, neurocognitive tests, clinical and diagnostic interviews, and self-reported childhood adversity. Follow-up assessments of functioning were conducted at 9 and 18 months. We applied multiblock sparse partial least squares (MB-SPLS), a multivariate data integration method that identifies latent structures of shared variation across modalities, within a leave-one-site-out cross-validation (LOSOCV) framework to ensure generalizability across sites. The derived latent variables (LVs) were examined for stability, interpretability, and biological plausibility. Functional outcome trajectories were modeled using longitudinal clustering of four functional domains (GAF symptoms, GAF disability, GF:Social, GF:Role), and machine learning models were trained using nested cross-validation to evaluate the predictive utility of the MB-SPLS-derived LVs. Clinical utility was further assessed via comparison with human rater predictions and decision curve analysis.<br>Results: We identified one robust and generalisable LV that loaded on neurodevelopmentally relevant features, including widespread cortical alterations, elevated polygenic risk scores for schizophrenia, bipolar disorder, and neuroticism, greater exposure to childhood trauma - particularly emotional abuse - and reduced cognitive performance, current functioning, and earlier social and scholastic premorbid adjustment. This multivariate profile was expressed across diagnostic groups but was most prominent in individuals with psychosis spectrum disorders (ICD-10 F2). The signature was significantly associated with poorer longitudinal functioning trajectories. Predictive modelling based on the latent scores revealed that genetic and neuroanatomical risk dimensions contributed most to distinguishing individuals with deteriorating functioning from those showing improvement - despite both groups being impaired at baseline. Compared to human predictions, the model demonstrated superior clinical utility, particularly in clinically challenging subgroups such as CHR. Decision curve analysis indicated superior net benefit of the multimodal model especially at higher threshold probabilities.<br>Conclusions and Relevance: The identified vulnerability signature reflects a transdiagnostic, dimensional neurodevelopmental risk profile, aligning with prior evidence on neurodevelopmental mechanisms in early-stage psychiatric disorders. This insight can inform stratified care: costly and resource-intensive assessments (e.g., MRI scans) may be most valuable in diagnostically ambiguous or clinically at-risk individuals, such as CHR cases. While MB-SPLS was not trained to predict functional outcomes, the derived LV proved clinically informative - bridging mechanistic understanding and actionable prognosis in precision psychiatry.

article VPE+26

Neuroscience Applied

5.106810. Jan. 2026.

Authors

C. S. Vetter • D. Popovic • F. F. Eichin • C. Weyer • A. Khuntia • K. Chrisholm • L. Kambeitz-Ilankovic • L. Antonucci • S. Ruhrmann • J. Kambeitz • M. Rosen • T. Lichtenstein • A. Riecher-Rössler • R. Upthegrove • R. K. R. Salokangas • J. Hietala • C. Pantelis • R. Lencer • E. Meisenzahl • S. J. Wood • P. Brambilla • S. Borgwardt • P. Falkai • A. Bertolino • D. Rückert • N. Koutsouleris