Using Machine Learning to Predict Features Within Substance Use Disorder Treatment Service Settings That Increase the Likelihood of Positive Treatment Outcomes.

Given the conceptual issues involved in defining and measuring recovery and accordingly substance use disorder (SUD) treatment outcomes, the role of each state's treatment system and social factors, the objective is to examine underlying and interrelated patterns within SUD treatment, outcomes, and recovery. Using a recovery-oriented framework, a Machine Learning Random Forest model was developed to analyze publicly funded SUD treatment services across the United States. The aim was to predict the 10 most important features that increase the likelihood of positive treatment outcomes, defined as less substance use (SU) or abstinence. Over 78% of SUD treatment services were provided to individuals either with Medicaid coverage or were uninsured. The most important feature identified was the number of days in treatment, regardless of setting. The second most important feature was the state and whether various treatment services were available. The third and fourth ranked features were the type of treatment at discharge and at admission, respectively. Housing status, SU self-help group participation, and employment were lower ranked. Referral source was the tenth ranked feature. The length of time in SUD treatment is consistent with the clinical perspective of the individual seeking treatment and continuing in care and recovery support. Individuals in Medicaid-funded treatment live in poverty, with peer support and community who have the least resources to support their recovery journey. States that prioritize behavioral health should coordinate to increase the availability of higher-cost, longer-duration treatment services across state lines, to states with low availability.