Decoding the adolescent non-suicidal self-injury: understanding with interpretable machine learning insights

Non-suicidal self-injury is a common risk behavior in adolescence but is often difficult to detect. This study employs interpretable machine learning techniques to develop a classification model for adolescent non-suicidal self-injury and elucidate pertinent factors. Employing diverse algorithms, a comprehensive analysis is conducted to discern critical risk and protective elements within a large dataset, evaluating their alignment with the Integrated Theoretical Model.

In partnership with educational institutions in eastern China, this research compiled data on behaviors and correlated factors through the administration of questionnaires, incorporating demographic information and seven validated scales. Analytical models were built using six machine learning techniques: K-Nearest Neighbors, Support Vector Machine, Logistic Regression, Light Gradient Boosting Machine, CatBoost, and eXtreme Gradient Boosting.

The analysis included a total of 2989 valid responses samples. Among the algorithms, CatBoost demonstrated superior performance, evidenced by an AUPRC of 0.736 and an AUC of 0.863. SHAP visualization highlighted 23 important items. Exploratory factor analysis identified seven factors, designated as Situational Anxiety, Depressive Symptoms, Positive Daily Functioning, Negative Self Esteem, Self-Appraisal of Behavior, Bullying and Reactive Aggression, and Interpersonal Problems and Self-Acceptance.

Leveraging multiple machine learning algorithms for a holistic item analysis, this research identifies critical risk and protective factors for non-suicidal self-injury, thus refining the Integrated Theoretical Model.