Predictive Modelling of Service Pathways to Admission in Psychiatric Residential Treatment Facilities

Objective To develop and test predictive models of admissions to a psychiatric residential treatment facility (PRTF) in transitional age youth using routinely collected health insurance claims and enrollment data. Data Sources We used outpatient service and pharmaceutical claims from Medicaid beneficiaries aged 6-to 21-years old in Kentucky for the years 2010-2017. Study Design We assessed over 1,250 predictors (derived from Medicaid claims data) prior to the first PRTF admission. An ensemble machine learning (ML) algorithm based on logistic regression models fitted to a random subsample of the original data was used to predict pathways to the first PRTF admission. Discrimination performance of the ML ensemble was evaluated by comparing predictions to actual outcomes and calculating area under the curve (AUC), accuracy, sensitivity, and specificity. Additionally, a multivariate logistic regression model was fit to investigate the contribution of the continuity of mental health care after the initial PRTF admission on the risk of readmission. Data Collection/Extraction Methods We identified N = 519,011 unique children and youth with at least one outpatient service or pharmaceutical claim during our study period (January 1, 2010 through December 31, 2017). Principal Findings Fewer than 0.5% of children and youth in Kentucky had an episode of PRTF admission. Despite a very low prevalence of PRTF admission, classification accuracy of the ML ensemble for identifying PRTF youth achieved over 90% accuracy (AUC = 0.96). Factors associated with the initial PRTF admission were having been prescribed anti-psychotic and anti-manic medications, and receiving outpatient psychiatric care. Within six months after the initial PRTF discharge, there was a surprising drop in service utilization with a large proportion of youth not appearing to receive any follow-up care. Conclusions Despite the fact that admission into a PRTF was a relatively rare event, our findings suggest that it is a predictable event among youth with identified mental health conditions who are receiving care in the community. What is known on this topic After psychiatric hospitalization, PRTF treatment is the most expensive and restrictive intervention available to serve children and youth. Previous research examining predictors of PRTF entry using Medicaid reimbursement data showed that clinical factors were strong predictors of hospitalization. What this study adds We provide a comprehensive analysis of the factors beyond clinical diagnoses that lead to PRTF entry. We also seek to identify whether any specific patterns of service and/or pharmacy claims utilization are associated with reducing the likelihood of readmission.