Using machine learning to predict heavy drinking during outpatient alcohol treatment

Area of Science:

• Psychiatry
• Data Science
• Clinical Research

Background:

• Accurate prediction is crucial for effective treatment of alcohol use disorder (AUD) and other psychiatric conditions.
• Traditional methods identified patient factors linked to AUD treatment outcomes.
• Limited research has systematically optimized predictive models for AUD treatment.

Purpose of the Study:

• To demonstrate machine learning's utility in predicting clinical outcomes for individuals undergoing outpatient AUD treatment.
• To develop and validate predictive models for key AUD treatment milestones.

Main Methods:

• Utilized data from the COMBINE multisite clinical trial (n=1383) for model development and testing.
• Employed the random forest algorithm to generate predictive models.
• Applied 'leave sites out' external validation and stratified analyses for sex differences.

Main Results:

• Models predicting first and last month heavy drinking achieved AUCs of 0.67-0.74 (internal) and 0.69-0.72 (external).
• The model for between-session heavy drinking demonstrated strong accuracy (AUC=0.89 internal, 0.80-0.87 external).
• Significant sex differences were observed in the optimal predictive feature sets.

Conclusions:

• Machine learning models can predict AUD treatment outcomes using routinely collected clinical data.
• This approach offers potential for significantly improved prediction accuracy without expensive assessments.
• Further research is necessary to optimize the deployment of these machine learning models in clinical settings.