JMIR Publications

ABSTRACT

Background:

The application of machine learning (ML) approaches in the diagnosis of alcohol use disorder (AUD) has expanded rapidly. However, the methodological rigor of existing models and their applicability in real-world clinical settings remain insufficiently characterized.

Objective:

To systematically evaluate the predictive performance, methodological quality, and clinical translation potential of ML-based models developed for AUD diagnosis and identification.

Methods:

PubMed, EMBASE, Web of Science, CNKI, Wanfang, VIP, and SinoMed databases were systematically searched from inception to July 2026. Original studies reporting the development or validation of predictive models for AUD diagnosis or identification were eligible for inclusion. The Prediction model Risk Of Bias Assessment Tool (PROBAST) was applied to evaluate risk of bias and applicability concerns. Random-effects models were used to pool the area under the curve (AUC), sensitivity, and specificity estimates.

Results:

A total of 33 studies involving 57 predictive models were included. The pooled AUC was 0.90 (95% CI, 0.87–0.92), with pooled sensitivity and specificity of 0.97 (95% CI, 0.93–0.99) and 0.98 (95% CI, 0.95–0.99), respectively. The pooled area under the summary receiver operating characteristic (SROC) curve was 0.99. Subgroup analyses demonstrated that deep learning models achieved a higher pooled AUC (0.944) than traditional machine learning approaches (0.893) and conventional statistical methods (0.869) (P < 0.05). Electroencephalography-derived features were the most frequently investigated predictors, followed by age and sex. All included studies were judged to have a high risk of bias, with the major methodological concerns being low events per variable (EPV < 20 in 13 studies), lack of external validation (19 studies), and inadequate handling of missing data.

Conclusions:

ML models demonstrate promising discriminative performance for AUD diagnosis; however, the current evidence base is substantially constrained by widespread methodological limitations. Future research should prioritize robust external validation, adherence to established methodological standards, and improved clinical usability to facilitate the translation of ML-based diagnostic tools into practice. Clinical Trial: This systematic review and meta-analysis was registered with PROSPERO (https://www.crd.york.ac.uk/PROSPERO/view/CRD420261484673)