Machine Learning-Based Computational Validation of the Addictions Neuroclinical Assessment Framework in relation to Hazardous Drinking

Background: Addiction is a multifaceted disorder driven by complex neurobiological and psychological mechanisms. The Addictions Neuroclinical Assessment (ANA) framework offers a dimensional mechanistic approach, focusing on three core domains: incentive salience, negative emotionality, and executive function. This study aimed to validate the ANA framework computationally in relation to hazardous drinking, with the hypothesis that incentive salience and negative emotionality would be most strongly associated with drinking severity. Methods: We analysed two independent datasets: a community cohort of 1,260 collected in 2016-2018 adults and a cohort of 655 young adult binge drinkers, collected between 2017 and 2018. The three ANA domains were operationalized using behavioural and self-report measures. Four machine learning models (elastic net, support vector machines, random forest, and gradient boosting machines) with nested 5-fold cross-validation were used to assess relations between ANA domains and hazardous drinking as measured via the Alcohol Use Disorder Identification Test (AUDIT), a validated screening instrument for hazardous drinking. Results: Across both datasets, elastic net consistently outperformed other models. Incentive salience, largely reflecting alcohol's reinforcing value, was most robustly related to AUDIT score (R² = .389–.419), followed by negative emotionality (R² = .293–.317), largely reflecting drinking to cope. Executive function, reflecting impulsivity and inhibitory control, accounted for less variance (R² = .098–.109). Optimizing elastic net models via meta-learner further improved performance, explaining more than half of the variance (R² = .539–.549). Limitations: These findings may not generalize to individuals who are older or have severe AUD. Cross-sectional data limits longitudinal causal inferences. Conclusion: These results provide robust computational validation for the ANA framework, emphasizing incentive salience and negative emotionality as key domains linked to AUDIT score. Future research should explore diagnostic and longitudinal applications.