Machine learning-based outcome prediction and novel hypotheses generation for substance use disorder treatment

Area of Science:

• Public Health
• Data Science
• Machine Learning

Background:

• Substance use disorder (SUD) is a significant public health challenge requiring effective treatment strategies.
• Understanding factors influencing treatment success is crucial for policy development.
• Traditional methods may overlook complex interaction effects.

Purpose of the Study:

• To propose a novel data analytics approach using machine learning to discover interaction effects impacting SUD treatment program success.
• To identify factors that may be neglected by traditional hypothesis-generating methods.

Main Methods:

• Joined patient-episode-level SUD treatment discharge data with FBI crime data.
• Applied machine learning models including random forests, artificial neural networks, and extreme gradient boosting with nested cross-validation.
• Identified and analyzed interaction effects using the best-performing model, followed by traditional logistic regression testing.

Main Results:

• Extreme gradient boosting achieved the highest performance (AUC 89.31%) in predicting treatment completion.
• Identified 73 potential interaction effects, with 12 proving statistically significant (P<.05) after logistic regression testing.
• Discovered novel interactions involving length of stay, substance use frequency, and self-help group attendance.

Conclusions:

• Novel interaction effects influencing SUD treatment completion were identified using a machine learning approach.
• These findings offer valuable insights for practitioners and policymakers to enhance treatment program effectiveness.
• The study highlights the utility of advanced data analytics in uncovering complex relationships in public health research.