Explainable AI on UK Biobank Data Ranks Sex, Tobacco Use and Brain Structure as Top AUD Predictors

Researchers applied multiple xAI techniques to more than 400 variables across five domains in the UK Biobank neuroimaging cohort. Case-control status was defined by AUD diagnoses and alcohol-use metrics. The model identified sex, lifetime tobacco smoking, cannabis use, age and regional brain volumes as the variables that most improved classification performance. Genetic ancestry and AUD polygenic scores also contributed, but socioeconomic status and several mental-health measures added little once overlapping biological and substance-use factors were accounted for. The analysis showed that standard linear models miss interactions such as those between sex and age or between social isolation and specific ancestry components. These interactions suggest distinct patient subtypes rather than uniform risk. Because the sample is restricted to European-ancestry individuals aged 40-69, the reported feature rankings cannot be assumed to generalize to younger cohorts or non-European populations without further testing. Future work requires replication in independent biobanks with longitudinal AUD incidence data and head-to-head comparison against conventional logistic regression and polygenic-risk scores alone. Only then can the xAI-derived subtypes be evaluated for clinical utility in prevention or treatment matching.