An Exploratory Study of Social Skills Deficits and Suicide Attempts in Adolescent Psychiatric Inpatients: A Machine Learning Analysis of Goldstein’s ART Framework

Background/Objectives: Social skills deficits are increasingly recognized as clinically relevant correlates of suicidal behavior in adolescents. Goldstein’s Aggression Replacement Training (ART) provides a structured framework for assessing discrete prosocial and emotional skills, but its relationship with suicidality in psychiatric inpatients has not been systematically examined. This paper investigated whether self-rated ART social skills items differentiate adolescents with and without a history of suicide attempt as well as whether machine learning (ML) models can identify interpersonal profiles associated with SA. Methods: In this cross-sectional study, 162 psychiatric inpatients aged 11–17 years (72% female) were assessed at admission. Social skills were measured using the 50-item Skillstreaming component of ART, which is grouped into six domains (basic, advanced, emotional, alternatives to aggression, stress management, planning). Machine learning models (logistic regression, random forest, XGBoost, SVM, k-NN, Gaussian Naive Bayes) were trained on item-level features using repeated nested cross-validation with a held-out test set providing an unbiased estimate of generalization. Results: One item—Helping Others—remained significant after Holm correction (adjusted p = 0.042) with higher scores among adolescents with a history of SA. Seven additional prosocial items reached nominal significance (raw p < 0.05) but did not survive correction. Across ML models, discrimination was modest (cross-validated ROC-AUC 0.59–0.66) with Gaussian Naive Bayes performing best (ROC-AUC 0.66). Held-out test set performance was lower (ROC-AUC 0.46, 95% CI 0.26–0.69). SHAP analysis localized the model’s signal to a coherent cluster of prosocial and empathic communication items with higher self-reported prosociality paradoxically associated with SA history. Conclusions: In this inpatient cohort, self-rated social skills carried a small but coherent discriminative signal for prior suicide attempt, which was concentrated in outward-directed prosocial behaviors rather than global interpersonal functioning. The counterintuitive directionality—higher prosocial self-ratings among attempters—aligns with developmental frameworks describing compliant, hyper-responsible, or self-silencing interpersonal styles. While predictive performance was modest and sample size limits generalizability, item-level ART assessment may generate hypotheses about interpersonal risk phenotypes in this specific population of predominantly female adolescent inpatients. Only one item survived correction for multiple comparisons, and all findings should be interpreted as exploratory and hypothesis generating.