Supplementary file 4_Can machine learning predict non-suicidal self-injury? A systematic r

Twelve studies involving 33,366 participants were synthesized to evaluate machine learning models for predicting non-suicidal self-injury. The review, published on figshare in 2026, found ensemble models showed a pooled AUC of 0.83, while single models had an AUC of 0.68. The evidence base is constrained by substantial heterogeneity and a high risk of bias in most studies.

Use Cases

• Benchmarking new machine learning models for NSSI prediction against the reported pooled AUC of 0.75.
• Informing study design for prospective NSSI prediction research based on identified limitations like cross-sectional data.
• Comparing the performance of ensemble versus single models for clinical risk assessment tasks.
• Assessing risk of bias in predictive modeling studies using the PROBAST+AI framework described.

Strengths

• Includes a quantitative synthesis of 12 studies and 33,366 participants.
• Provides specific performance metrics: pooled AUC of 0.83 for ensemble models and 0.68 for single models.
• Uses a formal risk-of-bias assessment tool (PROBAST+AI).
• Dataset is openly available under a CC-BY-4.0 license.

Limitations

• The dataset is a 245.6 KB document; it contains review text and results, not the underlying raw data from the included studies.
• Column-level documentation is absent; field semantics must be inferred after download.
• The review notes substantial heterogeneity and a high risk of bias in the analysis domain of most included studies.

Provenance

Source

figshare, authored by Qianhui Wen.

Collection Method

Systematic review and meta-analysis of studies from multiple databases searched from inception to June 28, 2025.

Time Range

Studies included from database inception to June 28, 2025.

Freshness

Last updated 2026-04-30 14:01:22.

Geography

null