Using machine learning to predict pretreatment drinking changes.

Area of Science:

• Addiction research
• Behavioral science
• Data science in healthcare

Background:

• Research indicates many individuals modify drinking behaviors before formal alcohol treatment.
• Previous studies on predictors of pretreatment drinking changes have yielded mixed results and small effect sizes.
• Theoretical and methodological limitations necessitate novel analytical approaches to understand these changes.

Purpose of the Study:

• To evaluate the predictive capabilities of traditional regression models and machine learning techniques for pretreatment drinking changes.
• To compare the performance of linear regression, logistic regression, recursive partitioning, random forests, neural networks, and support vector machines in predicting changes in drinking behavior.

Main Methods:

• Utilized baseline demographic and psychological data from 175 participants.
• Employed a training-testing split (80%/20%) for model development and validation.
• Developed models to predict percent change in drinking, heavy drinking days, and classify "pretreatment changers".

Main Results:

• Neural network models demonstrated the highest predictive accuracy, with areas under the curve ranging from poor to acceptable.
• Variable importance analysis identified demographic factors (education, income) and psychological constructs (processes of change) as significant predictors.
• Models predicted both continuous changes in drinking and classification of pretreatment changers.

Conclusions:

• Demographic variables are significant predictors of pretreatment drinking changes.
• Understanding societal and demographic factors is crucial for addressing alcohol use behaviors prior to treatment.
• Machine learning offers a promising avenue for analyzing complex patterns in addiction research.