University of Hawaiʻi at Mānoa applies Random Forest ML to 7.9M treatment records to identify substance use recovery predictors

Substance use disorder treatment outcomes were poorly understood at population scale. Traditional analytical methods could not detect patterns across millions of records or identify which treatment factors most strongly predict recovery, leaving policymakers without actionable guidance at a time when drug overdose deaths remain a major public health crisis.

Researchers at UH Center on Aging developed an ensemble Random Forest machine learning model to analyze 7.9 million publicly available U.S. treatment records. The model identified the 10 most important features predicting positive treatment outcomes and used ML visualization tools to map geographic disparities in treatment service availability across states.

Treatment duration emerged as the single strongest predictor of positive outcomes, regardless of setting. ML mapping revealed that states with the highest overdose death rates have fewer clinically appropriate treatment services — a pattern researchers said would have been 'virtually impossible to detect' without AI/ML. Findings were published in The Journal of Prevention Science and are informing state-level behavioral health policy recommendations.

• Length of time in treatment is the most critical modifiable factor for improving substance use disorder outcomes, regardless of treatment type or setting.
• AI/ML enables population-scale analysis of public health data that reveals geographic disparities invisible to traditional statistical methods.
• States with the greatest need (highest overdose rates) often have the least treatment infrastructure, pointing to a systemic equity gap that data-driven advocacy can help address.