Construction and internal temporal validation of a LASSO regression-based risk assessment model for non-suicidal self-injury addiction-like features in adolescents and young adults with depression.

Background: Non-suicidal self-injury (NSSI) is common among adolescents and young adults with depression and may present addiction-like features, increasing clinical complexity and suicide risk. However, clinically applicable tools for estimating the likelihood of NSSI addiction-like features remain limited. This study aimed to develop and internally temporally validate a LASSO-based cross-sectional risk assessment model for NSSI addiction-like features among patients with depression. Methods: A total of 511 adolescents and young adults with depression were included. All candidate predictors and the outcome were assessed at the same clinical evaluation; therefore, the model was interpreted as an association-based risk stratification tool rather than as evidence of causal or prospective prediction. Participants recruited during the first 3 months were assigned to the training cohort, and those recruited during the final month were assigned to the internal temporal validation cohort. LASSO regression was used for variable selection, and the selected variables were entered into multivariable logistic regression to construct the model and nomogram. Model performance was evaluated using ROC curves, calibration curves, calibration intercept and slope, Brier score, bootstrap validation, Hosmer-Lemeshow goodness-of-fit test, and decision curve analysis. Results: Among the 511 participants, 55.58% met the operational criterion for NSSI addiction-like features. LASSO regression identified age at first self-injury, gender, NSSI severity, borderline personality features, family functioning, and childhood trauma as the main associated variables. Older age at first self-injury, male gender, and better family functioning were associated with lower odds of NSSI addiction-like features, whereas greater NSSI severity, more prominent borderline personality features, and childhood trauma were associated with higher odds. The model yielded AUCs of 0.9366 in the training cohort and 0.9086 in the internal temporal validation cohort. The optimism-corrected AUC was 0.9310, and the more conservative bootstrap-corrected AUC was 0.9181. Brier scores were 0.1000 and 0.1220, respectively. In the validation cohort, the calibration intercept and slope were -0.1283 and 0.8895. Conclusion: NSSI addiction-like features were common in this clinical sample. The model showed promising discrimination, calibration, and potential clinical utility for cross-sectional risk stratification, but requires independent validation and prospective testing before routine clinical implementation.