From Risk Prediction to System Governance: A Scoping Review of Machine Learning Applications In Public Assistance and Social Service Systems

Background
Machine learning and predictive analytics are increasingly used in public assistance and social service systems to support targeting, service coordination, and administrative decision-making. However, prior reviews have focused primarily on individual-level risk prediction, especially in child welfare and mental health. System-level applications relevant to safety-net programs (e.g., TANF, SNAP, child care subsidies) and social services, such as demand forecasting, service capacity, and resource allocation, have received far less attention. Concerns also remain about limited reporting of calibration, subgroup performance, and external validation, all of which are essential for equitable and accountable use in public-sector settings. Guided by the Population–Concept–Context framework, this scoping review maps how predictive modeling has been applied across public assistance and social service systems, with attention to levels of prediction, evaluation practices, and policy translation.

Methods
This scoping review followed Arksey and O’Malley’s methodological framework and PRISMA-ScR reporting guidelines. A structured search identified 723 records across multiple databases, including PubMed/MEDLINE (n = 89), APA PsycINFO (n = 16), CINAHL (n = 28), Scopus (n = 340), and Web of Science (n = 118), supplemented by grey literature sources from Urban Institute and Child Trends. Eligible studies were empirical articles published in English between 2010 and 2026 that applied machine learning or predictive modeling to public assistance or social service settings using administrative or linked data. After screening, nine studies met inclusion criteria, highlighting the limited scope of the current evidence base. Included studies were coded by program domain, prediction level, data source, modeling method, outcome, evaluation approach, and policy application.

Results
Across 9 included studies, predictive modeling was applied to hospitalization risk, mental and behavioral health service utilization, long-term welfare dependence, SNAP recipient classification, need for wraparound services, long-term services and supports payment calibration, and Medicare Advantage network restrictiveness. Common approaches included supervised models, particularly logistic regression, random forests, and related classification methods. Most studies focused on individual-level risk stratification and service utilization, while relatively few examined system-level applications such as payment calibration, early warning systems, and network measurement. Where reported, model performance suggested meaningful administrative value. For example, one long-term services and supports payment model explained 21.6% of cost variation. Another study found that machine learning improved prediction of long-term welfare receipt by 22%, corresponding to a 14-percentage-point increase in R². Evaluation relied primarily on discrimination metrics such as AUC, whereas calibration, subgroup analysis, and external validation were reported consistently.

Conclusions
Predictive analytics in public assistance and social service systems remains concentrated on individual-level prediction, with limited attention to system-level, equity-oriented, and governance-focused applications. This review highlights gaps in evaluation transparency, fairness reporting, and real-world implementation. For social work research and practice, these results underscore the need to move beyond prediction toward accountable, policy-relevant uses of data-driven tools addressing demand forecasting, service access, and equitable resource allocation. By mapping this emerging field, this study provides a foundation for advancing more equitable and system-oriented applications of predictive analytics in publicly funded social programs and services.