A machine learning approach for identifying predictors of success in a Medicaid-funded, community-based behavioral health program using the Child and Adolescent Needs and Strengths (CANS)

Introduction: The CANS is the most popular measurement tool in the System of Care (SoC), with the potential to generate an estimated 1.9 million evaluations per year in the United States. This dataset has broad potential for decision support and outcomes monitoring, yet many SoC services do not yet leverage this information asset. We report here the results of a pilot project in which we applied machine learning methods to CANS data for the purpose of identifying clinical profiles associated with improvement in a public community-based behavioral health program in Pennsylvania. Methods: We analyzed over 7,000 CANS from 3,385 children who participated in Pennsylvania's Medicaid-funded behavioral health rehabilitation services (BHRS) program during 2013–2019. A gradient boosting classifier was developed to identify children most likely to experience a total score improvement on the CANS while participating in BHRS. Separate models were constructed for children with and without autism spectrum disorder (ASD). CANS-based clinical profiles associated with improvement were also identified. Analyses were run using Python Sci-Ki version 0.20.3 and Linearly Interpretable Model-agnostic Explanations (LIME) version 0.1.1.33. Results: The mean age of BHRS participants was 9.85 years (standard deviation: 3.47) and the majority were white (54%) or African American (11.5%). The median length of stay was 963 days (range: 541–2,008) and 39% (N = 1,330) had a diagnosis of ASD. A total of 49.9% of children had a CANS total score improvement. Precision of the gradient boosting classifier was 70% and 79% for children with and without ASD, respectively. Fourteen profiles were associated with improvement in ASD (mean probability of improvement per profile = 0.95, range: 0.88–1.0) and 55 such profiles were identified in children without ASD (mean probability of improvement: 0.91, range: 0.86–1.0). Conclusion: Machine learning can be applied to the CANS to identify children who have high probability of improvement in BHRS. These methods may have utility as an adjunct to existing decision support systems for SoC services.

Duke Scholars

Author Jesse David Troy Biostatistics & Bioinformatics, Division of Biostatistics