Improving Treatment Outcomes in People With Alcohol Use Disorder: Machine Learning Models Predict Heavy Drinking Relapse

Newswise — Machine learning techniques can be used to predict treatment outcomes among people with alcohol use disorder (AUD), according to a study reported in Alcoholism: Clinical and Experimental Research. Patients often return to heavy drinking during and after their treatment for AUD, and may need multiple rounds of treatment before they can achieve long-term abstinence from heavy alcohol use. The predictive models identified in this study could be used to develop clinical systems that allow clinicians and patients to anticipate drinking relapses and adjust treatment before they occur, enhancing clinical care and outcomes. Patient characteristics associated with relapse have been identified previously using traditional statistical techniques, but methodological limitations have made it difficult to leverage these associations to create optimal predictive models. Machine learning – a branch of artificial intelligence involving the use of computer algorithms that can build and automatically improve their inferences by analyzing patterns in sample data – is well suited to this task. In the current study, researchers from Yale University used machine learning to develop models to predict critical alcohol use outcomes in patients completing a structured outpatient treatment program, using routinely collected clinical data.

Data from a large clinical trial of interventions for AUD, the COMBINE trial, were used to develop and test the predictive models. Over 1300 adults with AUD had participated in the 16-week outpatient trial across 11 treatment centers. Patients were randomly assigned to one of nine different medication or behavioral therapy combinations, and underwent comprehensive and standardized evaluations at regular time points. This provided a large and rich dataset on which the machine learning algorithms could be “trained”. The goal was to develop a set of models capable of predicting heavy drinking (defined as 4/5+ drinks per day for women/men) at three different time points: during the first month of treatment, during the final month of treatment, and between weekly or bi-weekly treatment sessions.

The resulting models performed well in predicting the occurrence of heavy drinking across internal cross-validation data samples, as well as on data that had been deliberately excluded from the “training” dataset. This suggests that the models are generalizable across communities and providers. Individual predictive factors of most importance in the models included laboratory factors (such as liver enzyme levels), clinical factors (such as age of onset of alcohol dependence), and patient scores on self-report surveys (such as those relating to drinking behaviors and psychological symptoms) – all of which can be obtained relatively easily and inexpensively during treatment for AUD. The models showed substantial differences in the relative importance of specific predictive factors among men and women, consistent with previous research showing sex differences in factors linked to harmful alcohol use.

Compared with human clinical judgment, the machine learning models are likely to more accurately identify those patients at risk of returning to heavy drinking and who may therefore benefit from additional intervention during treatment. For example, patients who are most likely to drink heavily during the first or last month of treatment could be immediately diverted to a higher level of care, without incurring the financial and psychological cost of an unsuccessful treatment attempt. Similarly, identifying those most likely to drink heavily between treatment sessions could enable clinicians to preempt drinking lapses during high-risk periods. Future research could focus on the use of machine learning approaches to identify factors that predict patient responses to specific treatments.

Predicting heavy drinking during outpatient alcohol use treatment using machine learning. W. Roberts, Y. Zhao, T. Verplaetse, K. E. Moore, M. R. Peltier, C. Burke, Y. Zakiniaeiz, S. McKee (pages xxx)

ACER-21-5012.R2