Who is alcohol cue-reactive? A machine learning approach

Abstract

Background: The alcohol cue-exposure paradigm is widely used in alcohol use disorder (AUD) research. Individuals with AUD exhibit considerable variability in their alcohol cue-reactivity, highlighting the need to identify characteristics that contribute to this heterogeneity. This study applied machine learning models to identify clinical and sociodemographic predictors of subjective alcohol cue-reactivity (ALCUrge).

Methods: Individuals with AUD (N = 139; 83 M/56F) completed an alcohol cue-exposure paradigm and a battery of clinical and sociodemographic measures. ALCUrge (primary outcome variable) was assessed using the Alcohol Urge Questionnaire following alcohol cue-exposure. We implemented three machine learning models (Lasso regression, Ridge regression, Random Forest) to identify clinical and sociodemographic predictors of ALCUrge and compared model performance (i.e. predictive accuracy).

Results: Lasso regression had the strongest predictive accuracy, with a Root Mean Square Error (RMSE) of 9.48, followed by Random Forest (RMSE = 9.95), and Ridge regression (RMSE = 10.40). All models outperformed chance-level prediction (null baseline model RMSE = 14.80). Top predictors of ALCUrge across multiple models were alcohol urge prior to cue-exposure, compulsive alcohol-related behaviors/thoughts, tonic alcohol craving, cigarette smoking status, and biological sex. Higher pre-cue exposure alcohol urge, more compulsive alcohol-related tendencies, greater tonic craving, and occasional cigarette use was associated with greater predicted ALCUrge, while being female was associated with lower predicted ALCUrge.

Conclusion: This study advances our understanding of the phenotypic overlap in the compulsive aspects of tonic craving and phasic cue-induced alcohol urge, and offers insight into additional factors, such as biological sex and cigarette smoking, that may contribute to variability in alcohol cue-reactivity.

Keywords: addiction; alcohol use disorder; craving; cue-reactivity; machine learning; random forest; regression.

Figure 1

Frequency of alcohol-cue reactivity (ALC_Urge) scores. Frequency of AUQ total scores during the alcohol condition (ALC_Urge). On the AUQ, participants convey their present subjective urge to consume alcohol by reporting their agreement/disagreement with eight statements related to the subjective experience of alcohol urge on a seven-point Likert scale ranging from 0 (strongly disagree) to 6 (strongly agree). Total AUQ scores (possible range = 0 to 48) were calculated for each participant during the alcohol condition by summing responses to all eight items

Figure 2

Machine learning models: predictor variable importance. Variable importance for predictors in the (A) Lasso Regression, (B) Ridge Regression, and (C) Random Forest models. For Lasso and Ridge Regression, magnitude of coefficients indicates the importance of each variable. Negative coefficients indicate increased likelihood of reporting lower ALC_Urge, while positive coefficients indicate increased likelihood of reporting higher ALC_Urge. For Random Forest, variable importance was determined by mean decrease Gini. Gini impurity is a measure of node purity in decision trees. Mean decrease Gini explains which features are most influential in predicting the outcome. ^aOther Race: Includes participants who self-identified as Asian, Pacific Islander, American Indian, Alaska Native, mixed race, or a race other than White or Black RMSE: Root Mean Square Error; AUQ: Alcohol Urge Questionnaire; OCDS: Obsessive Compulsive Drinking Scale; ADS: Alcohol Dependence Scale; FTND1: Fagerstrom Test for Nicotine Dependence Question 1 (Never vs. Occasional vs. Daily Smoking); PACS: Penn Alcohol Craving Scale; TLFB: Timeline Followback; BDI: Beck Depression Inventory; BAI: Beck Anxiety Inventory