AI for Child Protection: Developing and Evaluating Machine Learning Applications for Decision Support in Child Protection

Description

Despite a high degree of uncertainty, child protection professionals must decide on appropriate interventions after determining that a child is at risk. While predictive risk models (PRMs) are used for risk assessment, there is a lack of evidence-based tools for selecting effective interventions. This project aims to use machine learning and artificial intelligence applications to develop decision support systems for child protection. The goal is to predict intervention effectiveness, ensure algorithmic fairness, and study the acceptance of these systems among professionals.
The project is divided into three phases:

1. Model Development: Development of predictive models to estimate the success of interventions based on administrative and unstructured case data. Classic regression methods and nonlinear classification models will be systematically compared.
2. Fairness Analysis: Development and application of statistical metrics to evaluate algorithmic fairness and identify and minimize biases in the models.
3. User Acceptance Analysis: Evaluation of three decision support prototypes — ranging from simple probability representations to reports generated by LLMs with visualizations using knowledge graphs — in A/B tests with experts.

This project is a collaboration with the Office for Youth and Vocational Guidance of the Canton of Zurich, utilizing state-of-the-art AI technologies, such as large language models (LLMs) and retrieval-augmented generation (RAG). The project contributes to the digital transformation of social work by developing practical, fair, and understandable tools for child protection.

Key data

Publications