Identifying predictors of treatment outcome in adults with alcohol use disorder
Background: Substance use disorders (AUD) affect health and wellbeing, and have broad societal costs. While many treatments exist, effectiveness is modest, leading to efforts to identify individual factors related to treatment success that might guide efforts to improve treatment outcomes. This approach has been tried, predominantly using theoretically defined individual characteristics (i.e., treatment groups for women), with limited improvements in outcomes. The present study used data-driven feature selection, enabling inclusion of a large number of baseline variables, to identify predictors of treatment success in an Intensive Outpatient Clinic setting.
Methods: Data from one year of successive admissions (2011 – 2012) to the Day One Intensive Outpatient Treatment Program at the University of Vermont Medical Center between was used to identify 140 individuals treated for alcohol use disorder. 102 variables were extracted through chart review and used in the analyses. First, 20% of the sample was set-aside for use in final model evaluation (and not used in model development). An Elastic Net Regularized linear regression, with 10-fold cross validation was used on the remaining 80% of the sample to develop the predictive model. The dependent variable was alcohol use (past 30 days) at the end of treatment.
Results: The study sample consisted of 140 subjects (mean age 38.3 ± 13.0; 37% female) who were predominantly Caucasian (97%) reflecting Vermont’s population demographics.
Subjects completed an average of 12 sessions (range 1 – 21) with 61% of participants reporting abstinence (30 day) at the end of treatment. Most subjects (65%) had a comorbid diagnosis (23% addition substance use disorder; 23% comorbid mental health disorder, and 19% both an additional substance use disorder and mental illness). 9 models were generated using the elastic net regularized regression, with models ranging in size from including 14 to 31 predictors and explaining between 26.4% – 34.5% of the variance in treatment outcome when tested in the independent set-aside sample of subjects. 11 variables appeared in 100% of the generated models and reflected baseline addiction severity, treatment characteristics (# of sessions attended, treatment compliance), and individual characteristics including baseline depression rating and readiness to change.
Conclusions: This study identified predictors of treatment outcome using a data-driven approach, which could be used to guide treatment efforts in the future. Interestingly, demographic variables that have been identified in previous research, using analytic methods that did not include cross-validation, were not replicated (e.g., age and gender), likely reflecting model overfitting in prior studies.