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

Area of Science:

• Artificial Intelligence in Medicine
• Medical Diagnostics
• Public Health Research

Background:

• Tuberculosis (TB) diagnosis relies on effective methods, with drug-resistant TB (DR-TB) posing a significant challenge.
• Machine learning (ML) shows potential for TB diagnosis, yet systematic evidence for DR-TB prediction and diagnosis is limited.
• Advancing artificial intelligence (AI) in healthcare requires robust evaluation of ML tools for complex diseases like DR-TB.

Purpose of the Study:

• To systematically review and meta-analyze the performance of ML models in diagnosing and predicting DR-TB.
• To consolidate evidence on ML efficacy for DR-TB, promoting AI adoption in this field.
• To compare diagnostic and predictive ML model performance and identify factors influencing accuracy.

Main Methods:

• A systematic literature search was conducted across major databases (PubMed, Cochrane, Embase, Web of Science) up to August 2025, supplemented by Google Scholar.
• Risk of bias was assessed using PROBAST, and a bivariate mixed-effects model was employed for meta-analysis of accuracy measures.
• Subgroup analyses were performed to stratify performance by ML task (diagnosis vs. prediction) and model type (e.g., deep learning vs. traditional ML).

Main Results:

• Twenty-six studies involving 35,472 participants were analyzed, revealing superior performance of diagnostic ML models (AUC 0.94) over predictive models (AUC 0.87).
• Deep learning (DL) diagnostic models achieved higher accuracy (AUC 0.97) compared to traditional ML (AUC 0.89).
• Internal validation demonstrated better performance than external validation for both diagnostic (AUC 0.95 vs. 0.85) and predictive models (AUC 0.88 vs. 0.85).

Conclusions:

• ML models, especially DL, are highly effective for DR-TB diagnosis, though their performance may decrease with external datasets.
• Predictive ML models offer moderate accuracy and are valuable for early risk stratification of DR-TB.
• Further large-scale, multi-center validation studies are essential to confirm the robustness and clinical utility of ML tools for DR-TB.