Classification of Alcohol Use Disorder With Explainable AI

The wide-ranging correlates (e.g., demographic, socioeconomic status [SES], brain, genetics) of alcohol use disorder (AUD) are typically characterized by small effects and studied in isolation. Novel multivariable nonlinear approaches may help characterize the multifaceted correlates of AUD and identify its strongest predictors.

Here, using Gradient Boosted Machines (GBMs), we classify AUD in the UK Biobank (N = 12,178) from features (n = 440) within the following 5 domains: (1) demographic (n = 12), (2) socioeconomic status (n = 18), (3) non-substance mental health (n = 55), (4) non-alcohol substance involvement (n = 6), and (5) biological features (n = 342). We use Explainable Artificial Intelligence (xAI) to characterize each respective feature's contribution to model performance and its heterogeneity, as well as the extent to which features from each domain collectively contribute to model performance.

In an out-of-sample hold-out test set (n = 6089), our model classified lifetime AUD well (AUC = 0.80). Notably, feature importance analyses revealed that demographic (age, sex, income), substance use (ever smoke tobacco or ever use cannabis), and mental health questionnaires (social isolation) were most informative to classifying AUD. The gradient boosted machine significantly outperformed a linear model and implicated non-zero pairwise interactions among variables, suggesting potential heterogeneity of AUD presentation by income and sex. Further, incrementally incorporating neuroimaging and genetic data as features led to small but consistent improvement in classifying AUD, with small effect sizes distributed across a number of variables (0%-2% gain based on overall model classification). This suggests the continued importance of biological features in classifying AUD beyond psychological and demographic measures, though classification with single selected brain regions may be limited.

Multimodal multivariable approaches considering nonlinear relationships may improve our understanding of AUD correlates. Particularly, accounting for interactions with demographic factors and integrating information across biological features may enhance model performance.