Related Experiment Video
Updated: Sep 12, 2025
Evidence-based Knowledge Synthesis and Hypothesis Validation: Navigating Biomedical Knowledge Bases via Explainable AI and Agentic Systems
Published on: June 13, 2025
Understanding the development, performance, fairness, and transparency of machine learning models used in child
Claudia Bull1, Steve Kisely2, Kim Betts3
1Princess Alexandra Hospital Southside Clinical Unit, Greater Brisbane Clinical School, Medical School, The University of Queensland, Queensland, Australia; Queensland Centre for Mental Health Research, The University of Queensland, Brisbane, Australia.
Objective:
To understand the development and validation of contemporary machine learning (ML) models for child protection prediction, their performance evaluation, integration of fairness, and operationalisation of model explainability and transparency.
Methods:
This systematic review followed the Preferred Reporting Items for Systematic reviews and Meta-Analyses (PRISMA) guidelines. Model transparency was assessed against the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis + Artificial Intelligence (TRIPOD+AI) criteria, while study risk of bias and model applicability were evaluated using Prediction model Risk Of Bias ASsessment Tool (PROBAST) criteria.
Results:
Eleven studies were identified, employing various ML approaches such as supervised classification models (e.g., binary classification, decision trees, support vector machines), regression models, and ensemble methods. These models utilised administrative health, child welfare, and criminal/court data. Performance was evaluated using a range of discrimination, classification, and calibration metrics, yielding variable results. Only four models incorporated group fairness, focusing on race/ethnicity as the protected attribute. Explainability and transparency were enhanced through Receiver Operating Curves, Precision-Recall Curves, feature importance plots, and SHapley Additive exPlanations (SHAP) plots. According to TRIPOD+AI criteria, only four studies reported likely reproducible models. Based on PROBAST criteria, all studies had unclear or high risk of bias.
Conclusions:
This is the first review to use TRIPOD+AI and PROBAST criteria to assess the risk of bias and transparency of ML models in child protection prediction. The findings reveal that the field remains methodologically immature, with many models lacking fair, transparent, and reproducible methods. Adoption of advanced fairness techniques (beyond fairness-through-unawareness), stakeholder involvement in model development and validation, and transparency through data and code sharing will be essential for the ethical and effective design of ML models, ultimately improving decision-making processes and outcomes for vulnerable children and families.
