BACKGROUND: Childhood trauma (CT) is a major yet elusive psychiatric risk factor, whose multidimensional conceptualization and heterogeneous effects on brain morphology might demand advanced mathematical modeling. Therefore, we present an unsupervised machine learning approach to characterize the clinical and neuroanatomical complexity of CT in a larger, transdiagnostic context.
METHODS: We used a multicenter European cohort of 1076 female and male individuals (discovery: n = 649; replication: n = 427) comprising young, minimally medicated patients with clinical high-risk states for psychosis; patients with recent-onset depression or psychosis; and healthy volunteers. We employed multivariate sparse partial least squares analysis to detect parsimonious associations between combinations of items from the Childhood Trauma Questionnaire and gray matter volume and tested their generalizability via nested cross-validation as well as via external validation. We investigated the associations of these CT signatures with state (functioning, depressivity, quality of life), trait (personality), and sociodemographic levels.
RESULTS: We discovered signatures of age-dependent sexual abuse and sex-dependent physical and sexual abuse, as well as emotional trauma, which projected onto gray matter volume patterns in prefronto-cerebellar, limbic, and sensory networks. These signatures were associated with predominantly impaired clinical state- and trait-level phenotypes, while pointing toward an interaction between sexual abuse, age, urbanicity, and education. We validated the clinical profiles for all three CT signatures in the replication sample.
CONCLUSIONS: Our results suggest distinct multilayered associations between partially age- and sex-dependent patterns of CT, distributed neuroanatomical networks, and clinical profiles. Hence, our study highlights how machine learning approaches can shape future, more fine-grained CT research.