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

Authors:

Abstract
Background: Child protection contacts are documented in the form of child protection intake reports detailing the child(ren) suspected to be at risk of harm, the notifier's concerns, the alleged person responsible, and other details. These narrative reports remain untapped for statistical reporting and research purposes due to being text-based as well as containing extremely sensitive content. Applying a contextual de-identification approach to child protection narratives opens up the possibility of large-scale automated analysis by external researchers while preserving data privacy and relational meaning. Data: 10,000 randomly selected South Australian Department for Child Protection intake reports were provided for this study. A total of 300 (200 development, 100 evaluation) were randomly selected, manually reviewed and annotated. Methods: We designed and implemented a context-preserving de-identification pipeline called De-Identification via Relationship-Based Replacement (DI-REBAR) on 10,000 reports while preserving the noted relationships between named individuals. A total of seven indirect identifiers (date, time, school, early learning centre, hospital, prison, sensitive reference numbers) and three direct identifiers (name, address, phone number) were selected for de-identification. A relationship table was temporarily created to map individuals and their relationship to the child of interest, with direct identifiers replaced. Indirect identifiers were replaced by list-based replacement and regular expressions. Results: DI-REBAR evaluation showed promising results in macro precision (>0.98), recall (>0.88), and F1-Score (>0.92) across nine out of the ten identifier classes (prison excluded). Each identifier had a range of precision (0.74-1.00), recall (0.61-1.00), and F1-Score (0.66-1.00). Address, phone number, and date have the best performances (>0.98) across precision, recall, and F1-Score. Conclusion: This is the first Australian study to automatically de-identify sensitive child protection data whilst preserving relationship information, allowing large-scale text analysis by external researchers in a data custodian's controlled environment. This approach can be applied to other areas that rely on unstructured data such as court documents and police reports. However, human reviewers are still needed to ensure identifiable information is not disclosed.

This publication has 21 references indexed in Scilit: