The quest for a biological phenotype of adolescent non-suicidal self-injury: a machine-learning approach

Abstract

AbstractNon-suicidal self-injury (NSSI) is a transdiagnostic psychiatric symptom with high prevalence and relevance in child and adolescent psychiatry. Therefore, it is of great interest to identify a biological phenotype associated with NSSI. The aim of the present study was to cross-sectionally investigate patterns of biological markers underlying NSSI and associated psychopathology in a sample of female patients and healthy controls. Comprehensive clinical data, saliva and blood samples, heart rate variability and pain sensitivity, were collected in n = 149 patients with NSSI and n = 40 healthy participants. Using machine-based learning, we tested the extent to which oxytocin, dehydroepiandrosterone (DHEA), beta-endorphin, free triiodothyronine (fT3), leukocytes, heart rate variability and pain sensitivity were able to classify participants regarding their clinical outcomes in NSSI, depression and borderline personality disorder symptomatology. We evaluated the predictive performance of several models (linear and logistic regression, elastic net regression, random forests, gradient boosted trees) using repeated cross-validation. With NSSI as an outcome variable, both logistic regression and machine learning models showed moderate predictive performance (Area under the Receiver Operating Characteristic Curve between 0.67 and 0.69). Predictors with the highest predictive power were low oxytocin (OR = 0.55; p = 0.002), low pain sensitivity (OR = 1.15; p = 0.021), and high leukocytes (OR = 1.67; p = 0.015). For the psychopathological outcome variables, i.e., depression and borderline personality disorder symptomatology, models including the biological variables performed not better than the null model. A combination of hormonal and inflammatory markers, as well as pain sensitivity, were able to discriminate between participants with and without NSSI disorder. Based on this dataset, however, complex machine learning models were not able to detect non-linear patterns of associations between the biological markers. These findings need replication and future research will reveal the extent to which the respective biomarkers are useful for longitudinal prediction of clinical outcomes or treatment response.