Adaptive Identification of Cortical and Subcortical Imaging Markers of Early Life Stress and Posttraumatic Stress Disorder

Abstract

Background and PurposePosttraumatic stress disorder (PTSD) is a heterogeneous condition associated with a wide range of brain imaging abnormalities. Early life stress (ELS) contributes to this heterogeneity, but we do not know how a history of ELS influences traditionally defined brain signatures of PTSD. Here we used a data-driven method to identify shared and unique neuroimaging markers of ELS and PTSD in 97 military veterans. We hypothesized that clustering standard regions of interest would improve classification accuracy of PTSD and ELS, relative to using individual atlas-defined brain regions, with moderate overlap in the patterns associated with each condition.\n\nMethodsWe used a novel machine learning method - evolving partitions to improve classification (EPIC) - to determine how combinations of cortical thickness, surface area, and subcortical brain volumes, could contribute to classification of veterans with PTSD (n=40) relative to trauma-exposed controls (n=57), and classification of ELS within the PTSD (ELS+ n=16; ELS- n=24) and control groups (ELS+ n=16; ELS- n=41). Additional inputs included intracranial volume, age, sex, and self-reported depression.\n\nResultsEPIC improved classification of PTSD, with 71% accuracy, on average. Within groups, EPIC classified ELS with 70% accuracy in controls on average, and 67% accuracy in PTSD; accuracy ranged from 75-84% for the best run of each analysis. For each binary analysis, EPIC identified unique predictors and combinations of neuroimaging variables that contributed to improve group classification from the classifier that modeled input features individually. Self-reported depression was a stronger predictor of PTSD and ELS than any combination of neuroimaging variables, along with adult trauma exposure in controls. Temporoparietal metrics were the strongest neuroimaging predictors of PTSD after depression, whereas regions of the cingulate cortex were strong markers of ELS in both subgroups.\n\nConclusionsCombinatorial machine learning methods such as EPIC may boost power to delineate underlying structural signatures of stress-related neuropathology.