Using Machine Learning To Predict Future Foster Care Admission

Structured Abstract Introduction Foster care admissions are highly traumatic for children and their families, often causing serious adverse outcomes. We seek to assess the viability of machine learning methods to identify children at risk of future foster care admission to facilitate diversion. Methods We use claims data for children enrolled in a Medicaid health plan in Ohio as well as for linked adults, along with data on individual and geographic Social Determinants of Health (SDOH) factors. We test the performance of a gradient-boosted tree machine learning algorithm as compared to logistic regression. 85% of children have SDOH data available. Results Using a gradient-boosted tree machine learning algorithm, we built a model that identifies 2,408 children (1.32%) as at risk of foster care admission in a sample of 181,841, of whom 1,599 entered foster care within one year, resulting in a positive predictive value PPV of 66.4% (F1=55.5%, Specificity=99.5%, Sensitivity=47.67%), outperforming logistic regression. Accuracy was substantially better when using SDOH data (PPV of 84.72% with SDOH data compared to 27.44% without). Conclusions These results highlight the importance of SDOH factors in predicting foster care admission. They also point to the potential of machine learning for facilitating early intervention to prevent foster care admissions.