Identification of hazardous alcohol use in outpatient psychiatric care: A comparison of biomarker phosphatidylethanol (PEth) and self-report.

Abstract

Phosphatidylethanol (PEth) is a reliable biomarker of recent hazardous alcohol use but is resource-intensive and not always feasible in routine mental health care. The Alcohol Use Disorders Identification Test (AUDIT-10) and its short form, AUDIT-C, are widely used self-report tools that may offer practical alternatives. This study examined whether AUDIT scores can serve as proxies for PEth and whether machine learning models enhance prediction. Data were collected from 4063 psychiatric outpatients who completed both PEth testing and AUDIT assessments. Regression models evaluated associations between AUDIT scores and PEth concentrations. Receiver Operating Characteristic (ROC) analyses determined cut-offs for hazardous alcohol use (PEth ≥ 0.3 µmol/L). Logistic regression, random forest, and XGBoost models were trained using AUDIT-10/AUDIT-C scores, age, and gender. AUDIT-10 and AUDIT-C correlated strongly with PEth (pseudo-R² = 30-47 %). ROC analyses showed good discrimination for hazardous use: AUDIT-10 AUC = 0.80 (optimal cut-off ≥5) and AUDIT-C AUC = 0.83 (optimal cut-off ≥4). XGBoost models improved classification modestly, yielding AUCs of 0.90 (AUDIT-10) and 0.88 (AUDIT-C), with balanced accuracies of 79-83 %, outperforming logistic regression and random forest. AUDIT-10 and AUDIT-C are effective, accessible tools for identifying hazardous alcohol use in psychiatric populations, supporting their role as practical alternatives when PEth testing is unavailable. Machine learning methods offer incremental gains, but simple cut-off scores remain clinically useful. Findings highlight the value of integrating brief alcohol screening into psychiatric care to guide timely interventions, while considering PEth's biological variability.