Automated de-identification preserves relationships in child protection reports

An automated pipeline developed to de-identify child protection intake reports while retaining the relationships described in their narratives achieved macro precision above 0.98, recall above 0.88 and an F1 score above 0.92 across nine of ten identifier classes in an evaluation of 10,000 South Australian Department for Child Protection reports. The method, called De-Identification via Relationship-Based Replacement, or DI-REBAR, replaced direct and indirect identifiers rather than simply removing sensitive text, preserving information about how people appearing in each report were connected to the child of interest. The prison category was excluded from the reported macro scores, and the study’s authors emphasized that human review remains necessary to check that identifying information has not been disclosed.

The study examined narrative intake reports documenting children suspected to be at risk of harm, a notifier’s concerns, people alleged to be responsible and other details. Because these records are text-based and contain extremely sensitive content, they have remained largely unavailable for statistical reporting and research. Researchers randomly selected 10,000 reports from the South Australian department for the project.

Of these, 300 were randomly selected for manual review and annotation, with 200 assigned to development of the pipeline and 100 to evaluation. The research therefore tested a de-identification approach on a defined collection of South Australian child protection intake reports rather than on live operational use or on the likely outcomes of downstream analysis. DI-REBAR targeted three direct identifier classes—names, addresses and phone numbers—and seven indirect identifier classes comprising dates, times, schools, early learning centres, hospitals, prisons and sensitive reference numbers.

The method created a temporary relationship table mapping individuals to their recorded relationship with the child of interest. Direct identifiers were replaced within that structure, allowing the pipeline to retain relational meaning while changing identifying details. Indirect identifiers were handled with list-based replacement and regular expressions.

Evaluation showed that performance differed among the identifier types. Across all ten classes, precision ranged from 0.74 to 1.00, recall from 0.61 to 1.00 and F1 scores from 0.66 to 1.00. Address, phone number and date produced the strongest reported results, each exceeding 0.98 for precision, recall and F1 score.

The authors reported the macro precision, recall and F1 figures above 0.98, 0.88 and 0.92, respectively, for the nine classes included in that summary; the prison category was not included in those macro scores. These results describe performance on the manually annotated evaluation sample and do not establish that all identifying details were removed from every one of the 10,000 reports. The authors described DI-REBAR as the first Australian study to automate the de-identification of sensitive child protection data while preserving relationship information.

They said this could make large-scale analysis of child protection narratives possible for external researchers working in a data custodian’s controlled environment, while protecting privacy and retaining relationships among the people described. They also identified other potential uses for similarly structured data, including court documents and police reports, but the reported evidence concerns the child protection intake-report evaluation. Human reviewers are still required to ensure that identifiable information is not disclosed, and the study did not test whether de-identified reports improve child protection decisions or whether the pipeline is safe and effective during live operational use.

Subject of Research: Evaluation of automated relationship-preserving de-identification of child protection intake reports

Article Title: De-identification of child protection intake reports using relationship-based replacement (DI-REBAR)

Article References: Lukmanjaya, W., Bulter, T., Perez-Concha, O., Cox, S., Bromfield, L., & Karystianis, G. (2026). De-Identification of child protection intake reports using Relationship-Based Replacement (DI-REBAR). Frontiers in Digital Health, 8, 1949307. https://doi.org/10.3389/fdgth.2026.1949307

Image Credits: AI Generated

DOI: 10.3389/fdgth.2026.1949307

Keywords: child protection, de-identification, natural language processing, privacy, pseudonymization, text analysis, sensitive data, relationship preservation, South Australia, child, protection, intake