Decoding rumination: A machine learning approach to a transdiagnostic sample of outpatients with anxiety, mood and psychotic disorders.

OBJECTIVE: To employ machine learning algorithms to examine patterns of rumination from RDoC perspective and to determine which variables predict high levels of maladaptive rumination across a transdiagnostic sample. METHOD: Sample of 200 consecutive, consenting outpatient referrals with clinical diagnoses of schizophrenia, schizoaffective, bipolar, depression, anxiety disorders, obsessive compulsive and post-traumatic stress. Machine learning algorithms used a range of variables including sociodemographics, serum levels of immune markers (IL-6, IL-1ß, IL-10, TNF-α and CCL11) and BDNF, psychiatric symptoms and disorders, history of suicide and hospitalizations, functionality, medication use and comorbidities. RESULTS: The best model (with recursive feature elimination) included the following variables: socioeconomic status, illness severity, worry, generalized anxiety and depressive symptoms, and current diagnosis of panic disorder. Linear support vector machine learning differentiated individuals with high levels of rumination from those ones with low (AUC = 0.83, sensitivity = 75, specificity = 71). CONCLUSIONS: Rumination is known to be associated with poor prognosis in mental health. This study suggests that rumination is a maladaptive coping style associated not only with worry, distress and illness severity, but also with socioeconomic status. Also, rumination demonstrated a specific association with panic disorder.