(PS9-D79) Comparison of Six Regression-based Machine Learning Algorithms to Predict NSSI Severity

Non-suicidal self-injury (NSSI) is a global public health concern. NSSI is the deliberate and intentional injury to body tissue without suicidal intent (Hooley et al., 2020) and is associated with increased risk for mental health problems. Understanding how to best intervene with individuals who engage in NSSI can ease the long-term negative (sometimes life-threatening) consequences of NSSI. The present study aims to address this pressing public health issue by contributing to the accurate assessment of NSSI risk, using a machine learning (ML) approach to predict NSSI severity. Data were collected online from a sample of undergraduate students (
N

= 863). A total of six regression-based ML algorithms were used to model NSSI severity. A supervised ML approach was used to estimate NSSI severity. First, data were randomly allocated into a training set (70%;
n

= 604) and a test set (30%;
n

= 259; a subset of the data reserved for testing of the model). Repeated cross-validation was used to iteratively train the models on the training set. After training, the models were tested on the test set. Predictor variables were selected based on the Cognitive-Emotional Model of NSSI and empirical evidence for risk factors of NSSI. Our results indicated that shrinkage algorithms (lasso, elastic net, and ridge regression) performed best at predicting NSSI severity using the selected predictors. Additionally, across all three algorithms, symptoms of borderline personality disorder (BPD), exposure to traumatic events, and depression symptoms respectively had the highest relative importance values for predicting NSSI severity. Clinical implications include the use of measures of BPD, trauma history and depression symptoms to inform NSSI risk assessment.