Multimodal phenotypic classification of generalized anxiety and panic using structural MRI data and psychosocial factors: Machine learning results from the German National Cohort (NAKO) study

Anxiety disorders (ANX) are common and impairing mental health conditions. This study aimed to classify self-reported symptoms of generalized anxiety disorder (GAD) and panic attacks as two psychopathological manifestations of ANX by applying machine learning to a cross-sectional dataset of 26,378 adults from the German National Cohort Study (NAKO). We first explored linear relationships between preselected neuroimaging correlates in MRI scans and anxiety phenotypes. Overall, sex-stratified correlation coefficients - while partly highly significant - were extremely low with r ≤ .04 for panic attacks and r ≤ .06 for GAD symptoms after correction for confounding variables like childhood trauma and depression. We then examined the combined classifying value of whole-brain imaging data of 246 ROIs in addition to psychosocial variables such as self-reported depression symptoms, stress, and childhood trauma, using four machine learning algorithms (support-vector machines with linear and radial kernels, elastic-net regression, and random forest). Neuroimaging data, particularly gray-matter volumes in regions such as the amygdala and superior parietal lobule, contributed to classification, but performance was substantially better when psychosocial variables were added. For both GAD symptoms and panic attacks, depression, stress and childhood trauma were the clearest indicators the classification would show the condition was present. Random forest models based on psychosocial variables alone achieved the highest discrimination performance for GAD symptoms (area under the receiver operating characteristic curve, AUROC = 0.973) and panic attacks (AUROC = 0.933). Combining neuroimaging and psychosocial variables in elastic-net regressions further improved specificity. These results support multimodal approaches to diagnose and investigate ANX that integrate structural brain abnormalities and psychosocial measures to capture the complexity of GAD and panic attacks, enabling the creation of individual risk profiles based on multiple biomarkers. These profiles may guide tailored therapeutic and preventive interventions.