Can machine learning predict non-suicidal self-injury? A systematic review and meta-analysis

AI Summary

Ensemble ML models demonstrated superior discrimination (pooled AUC 0.83; sensitivity 0.78; specificity 0.73), outperforming single models (AUC 0.68).
Overall evidence limited by substantial heterogeneity, high analysis-domain risk of bias, and predominance of cross-sectional designs hindering prospective prediction.
Translation to practice premature; prospective multicentre studies, external validation, and standardised reporting are required before clinical or public health use.

Front Public Health. 2026 Apr 30;14:1763121. doi: 10.3389/fpubh.2026.1763121. eCollection 2026.

ABSTRACT

BACKGROUND: Non-suicidal self-injury (NSSI) is common among adolescents and young adults and remains difficult to detect early using conventional approaches. machine learning (ML) has increasingly been applied to develop prediction models for NSSI.

METHODS: We conducted a systematic review and meta-analysis of studies that developed ML models for NSSI prediction, as defined by the original study authors. Multiple databases were searched from inception to June 28, 2025. Model performance, including the area under the curve (AUC), sensitivity, and specificity, was synthesized using a bivariate random-effects model. Risk of bias was assessed using PROBAST+AI.

RESULTS: Twelve studies involving 33,366 participants were included. In the primary model-level analysis, ensemble models showed relatively favorable pooled discrimination, with a pooled AUC of 0.83 (95% CI: 0.79-0.86), sensitivity of 0.78 (95% CI: 0.68-0.85), and specificity of 0.73 (95% CI: 0.58-0.84). Single models showed lower performance (AUC: 0.68, 95% CI: 0.64-0.72). Only one study evaluated a deep learning (DL) model (AUC = 0.70), and this estimate should therefore be interpreted cautiously. Across all 19 models, the pooled AUC was 0.75 (95% CI: 0.71-0.79). Substantial heterogeneity was observed, and the apparent advantage of ensemble models was not sustained in the study-level sensitivity analysis. Most studies were judged to be at high risk of bias in the analysis domain.

CONCLUSIONS: ML models show promise for identifying NSSI-related risk, but current evidence supporting true prospective prediction remains limited. The evidence base is constrained by substantial heterogeneity, a high risk of bias, and the predominance of cross-sectional studies. Prospective multicenter studies with external validation and standardized reporting are needed before ML-based models can be translated into clinical or public health practice.

SYSTEMATIC REVIEW REGISTRATION: https://www.crd.york.ac.uk/PROSPERO/view/CRD420251075613, identifier: CRD420251075613.

PMID:42145491 | PMC:PMC13173908 | DOI:10.3389/fpubh.2026.1763121

Document this CPD