Abstract

Background: Prediction models for child maltreatment risk are increasingly used to support decisions in child protection, yet concerns remain about methodological quality, transparency, calibration, and equity, particularly when tools are derived from administrative data.

Objective: To systematically review prediction models for child maltreatment risk and evaluate adherence to TRIPOD, risk of bias and applicability using PROBAST, and the extent of evidence on calibration, external validation, and fairness.

Methods: We included quantitative studies that developed or validated multivariable prediction models for maltreatment-related outcomes in child protection or public health contexts. Electronic databases and registers (2010-2025) were searched for studies reporting model performance. Two reviewers independently screened records, extracted data, and appraised reporting using TRIPOD and risk of bias/applicability using PROBAST. Owing to heterogeneity in outcomes, model types, and data sources, findings were synthesized narratively.

Results: Fourteen studies met inclusion criteria. Most used administrative or clinical datasets and logistic regression or machine learning models, achieving moderate to high discrimination. Five themes emerged: partial TRIPOD adherence; frequent analysis-domain bias; limited calibration and decision-analytic evaluation; sparse external validation and model updating; and uneven fairness auditing.

Conclusions: Current maltreatment prediction models show promising discrimination but are constrained by incomplete reporting, methodological weaknesses, and limited evidence on calibration, transportability, and equity. Future work should align with TRIPOD and PROBAST, embed validation and calibration, and incorporate fairness auditing.