Accuracy of Machine Learning in Identifying Drug Resistance in Tuberculosis: A Systematic Review and Meta-Analysis

Background and aims Machine learning (ML) has shown promise in diagnosing tuberculosis (TB), but systematic evidence on its role in predicting and diagnosing drug-resistant tuberculosis (DR-TB) is lacking. This study integrates a systematic review and meta-analysis to consolidate ML's performance in DR-TB diagnosis and prediction to promote artificial intelligence in this field. Methods Relevant studies were retrieved from PubMed, Cochrane, Embase, and Web of Science up to August 20, 2025, complemented by a manual search of Google Scholar. Risk of bias was evaluated with PROBAST. A bivariate mixed-effects model pooled accuracy measures, with subgroup analyses stratified by ML tasks (diagnosis and prediction). Results Twenty-six studies, including 35,472 participants, were analysed. Diagnostic models outperformed prediction models, with a higher pooled AUC (0.94 vs. 0.87). Deep learning (DL)-based diagnostic models consistently surpassed traditional ML across all key metrics, AUC (0.97 vs. 0.89). In the diagnostic model, internal validation showed superior performance to external validation AUC (0.95 vs. 0.85), and in the predictive model, the overall performance of the model in internal validation is slightly better than that in external validation AUC (0.88 vs. 0.85). Conclusion ML models, particularly DL, demonstrate high diagnostic efficacy for DR-TB, though performance declines in external data sets. Predictive models show moderate accuracy but remain useful for early risk stratification. Large multi-center validations are needed to ensure robustness and clinical applicability.