Using Machine Learning to Classify Individuals With Alcohol Use Disorder Based on Treatment Seeking Status

Area of Science:

• Neuroscience
• Psychiatry
• Machine Learning in Healthcare

Background:

• Less than 10% of individuals with Alcohol Use Disorder (AUD) receive treatment, indicating a significant gap in care.
• Understanding factors associated with treatment utilization is crucial for both clinical practice and research to improve AUD outcomes.
• Phenotypic differences between treatment-seeking and non-treatment-seeking individuals can confound research findings and impact clinical approaches.

Purpose of the Study:

• To identify a minimal set of clinical measures that accurately predict treatment-seeking behavior in individuals with AUD.
• To leverage machine learning, specifically alternating decision trees (ADT), for deep phenotyping in AUD research.
• To provide clinically relevant insights that can inform targeted treatment strategies and improve AUD research.

Main Methods:

• An alternating decision tree (ADT) classifier was constructed using 178 clinical measures from 778 individuals with AUD.
• The ADT model identified a subset of 10 measures that best predicted treatment-seeking status.
• Model validation was performed using cross-validation and an independent dataset of 236 individuals.

Main Results:

• The 10 key predictors included drinking behavior, depression, psychological problems, and substance dependence.
• The ADT achieved 86% accuracy in cross-validation and 78% accuracy on an independent dataset.
• The ADT model required fewer measures than a comparable logistic model to achieve similar predictive accuracy.

Conclusions:

• A subset of 10 clinically relevant measures can accurately classify individuals with AUD based on treatment-seeking status.
• Drinking behavior and depression emerged as strong predictors of treatment seeking.
• These identified measures can serve as potential targets for intervention and inform clinical research on AUD treatment.