Machine Learning for Biomarker-Based Tuberculosis Diagnosis: A Systematic Review and Meta-Analysis

Abstrak

Tuberculosis (TB) is a serious worldwide health concern, demanding reliable and efficient diagnostic techniques. Machine learning (ML) techniques have emerged as viable ways to enhance tuberculosis detection through the analysis of complicated biomarker data. However, the comparative effectiveness of various ML models is unclear, emphasizing the need for a thorough assessment. This study evaluates the diagnostic accuracy of ML algorithms for tuberculosis detection, outlining their strengths, limitations, and clinical applications. A comprehensive search of 6 databases identified studies using biomarker-based ML approaches for TB diagnosis. A 2-level mixed-effects logistic regression model was used to assess pooled sensitivity, specificity, and AUC. The RoB 2.0 tool was used to evaluate quality, whereas RevMan 5.4 was used for meta-analysis. Among the algorithms tested, the Probabilistic Neural Network (PNN) had the greatest pooled sensitivity (96.1%), specificity (89.9%), and AUC (0.94). Decision Tree (DT) models have great sensitivity (95.2%), but poor specificity (58.7%). Naïve Bayes (NB) had the greatest specificity (91.5%) and sensitivity (79.6%), while Random Forest (RF) and Support Vector Machine (SVM) performed similarly, with sensitivities of 83.7% and 84.1%, and accuracies of 82.5% and 83.8%, respectively. Logistic regression (LR) had the lowest sensitivity (54.3%) and accuracy (73.1%). ML algorithms have great potential for improving TB diagnosis using biomarker data, with PNN appearing as the best-performing model in terms of sensitivity, specificity, and AUC. However, the availability and expense of biomarker testing may limit the use of such strategies in resource-constrained environments. The clinical context, diagnostic goals, and infrastructure capabilities should all be considered while selecting algorithms. HIGHLIGHTS Machine learning (ML) algorithms show strong potential for biomarker-based tuberculosis (TB) diagnosis. The Probabilistic Neural Network (PNN) achieved the highest diagnostic accuracy among all models. Decision Tree and Naïve Bayes demonstrated strengths in sensitivity and specificity, respectively. Random Forest and Support Vector Machine offered balanced diagnostic performance. ML-assisted biomarker analysis could enhance TB detection, especially when optimized for clinical context. GRAPHICAL ABSTRACT

Identifier

• OpenAlex: https://openalex.org/W4415547086
• DOI: https://doi.org/10.48048/tis.2026.11390