Identification of distinct clinical phenotypes and their neurobiological signatures in stress-exposed individuals: a multimodal machine learning approach

Background: Individual responses to stress are highly heterogeneous, resulting in diverse psychopathological outcomes.This variability poses challenges for traditional diagnostic frameworks and underscores the need for a transdiagnostic approach to guide effective interventions.This study aimed to identify distinct phenotypes within a stress-exposed population and to characterize their associated biological profiles using a multimodal machine learning framework.Methods: A total of 809 stress-exposed adults (mean age 40.5 ± 8.74 years; 53.7% female) underwent comprehensive clinical, laboratory, and structural MRI assessments.Data-driven clustering of clinical variables was applied to identify phenotypes, followed by the development of machine learning classifiers trained on neuroimaging and laboratory data to predict phenotype membership.SHAP (SHapley Additive exPlanations) analysis was used to identify key biological features distinguishing each phenotype.Results: Three phenotypes were identified: a 'multi-risk' group (n=321) characterized by prominent depression, anxiety, and sleep disturbances; an 'alcohol-related risk' group (n=226) with high alcohol misuse and minimal comorbidity; and a resilient 'low-risk' group (n=262).Machine learning models accurately classified these phenotypes, indicating distinct underlying biological profiles.SHAP analysis revealed phenotype-specific signatures: the multi-risk phenotype was associated with frontal-subcortical structural alterations and dysregulated cortisol, whereas the alcohol-related risk phenotype was characterized by frontal-insular structural alterations and metabolic abnormalities.Conclusions: This study demonstrates the stratification of stress-exposed individuals into clinically and biologically distinct phenotypes.By integrating multimodal data with machine learning, we identified phenotype-specific neurobiological and metabolic profiles that extend beyond conventional diagnostic frameworks.These findings support a transdiagnostic, data-