Artificial Intelligence and Child Abuse and Neglect: A Systematic Review

Child maltreatment remains a global concern requiring urgent attention due to the severe long-term health consequences for victims. While digital technologies offer potential solutions, the current landscape of automated detection tools remains unclear. No prior work had resolved the extent to which machine learning supports protective services. That uncertainty drove the need for a comprehensive assessment of existing predictive frameworks. Researchers have yet to clarify how these advanced algorithms perform when applied to sensitive pediatric datasets. This gap motivated an investigation into the maturity of current computational implementations. Existing literature lacked a standardized evaluation of methodological quality across these diverse technological approaches. Consequently, the field currently operates without a clear understanding of the reliability of these automated systems.

The aim of this study was to determine the current state of development and validation for predictive models used to address child abuse and neglect. Researchers sought to clarify whether these automated tools provide reliable support for protective services. The investigation specifically addressed the lack of knowledge regarding the maturity of these computational implementations. By conducting a systematic review, the authors intended to identify the types of models currently in use. Another primary objective was to evaluate the methodological quality and risk of bias within the existing literature. This work was motivated by the absence of comprehensive summaries regarding these digital interventions. The authors also wanted to identify specific shortcomings in how these studies manage data and validation. Ultimately, the study provides a necessary overview of the current evidence to inform future research priorities.

The review approach involved a systematic search of the PubMed database to identify relevant literature. Investigators established strict inclusion criteria for articles written in English between January 1985 and March 2023. The team required that all selected publications utilize medical or protective service datasets for their analysis. Reviewers performed a comprehensive screening process that initially yielded 413 potential records. From this initial pool, the researchers selected seven papers that met all predefined requirements for inclusion. The team then extracted data regarding the types of models and the specific input information used in each study. They assessed the methodological quality of each paper to determine the overall risk of bias. This structured design allowed for a focused examination of the current state of predictive modeling in this field.

Key findings from the literature reveal that the application of these models remains in a nascent stage. The analysis identified seven studies that met the criteria for inclusion in the final review. These papers utilized diverse input data types to train their predictive systems. The researchers observed that the median size of the datasets employed was 2600 cases. Common computational approaches included natural language processing, convolutional neural networks, and standard artificial neural networks. Every single study included in the review demonstrated a high risk of bias. The results indicate that the field is significantly behind other medical sectors in terms of technological maturity. These findings underscore a lack of consistency in how researchers develop and validate their predictive frameworks.

The authors suggest that the adoption of automated predictive tools in this sector trails behind other clinical domains. Synthesis and implications indicate that current research suffers from consistent methodological shortcomings. Every included investigation displayed a significant risk of bias, undermining the robustness of their findings. The researchers propose that future efforts prioritize rigorous validation protocols to ensure model accuracy. Proper management of missing information and overfitting is necessary to improve the quality of future evidence. Investigators should also focus on selecting appropriate sample sizes to enhance the generalizability of their results. These findings imply that the field requires a more standardized approach to model development and reporting. Overall, the evidence highlights a need for greater caution and improved scientific standards in this sensitive area of research.