Integrative machine learning approaches with genomic data for predicting antitubercular drug resistance: A systematic review and meta-analysis.

OBJECTIVES

Tuberculosis (TB) remains a leading infectious cause of death, with drug-resistant TB threatening control gains worldwide Rapid, accurate prediction of resistance to first-line agents is essential to guide therapy. The objective of this meta-analysis is to evaluate the diagnostic performance of machine learning (ML) algorithms trained on genomic data for predicting phenotypic resistance to the antitubercular drugs.

METHODS

We searched 4 databases for original studies applying ML to whole-genome or targeted sequencing for resistance prediction against rifampicin (RIF), isoniazid (INH), ethambutol (EMB), streptomycin (STM), and other routinely tested drugs. Random-effects meta-analyses (REML) pooled sensitivity and specificity; Area Under Curve (AUC) was meta-analysed after logit transformation. Publication bias was examined via funnel plots plus Egger's and Begg's tests with trim-and-fill adjustment.

RESULTS

Seven eligible studies were included. Pooled performance was strongest for RIF and INH (RIF: sensitivity 0.90, specificity 0.95; INH: sensitivity 0.88, specificity 0.93). EMB and STM showed lower sensitivities despite reasonable AUCs. Forest plots for AUC (including logit scale), sensitivity, and specificity demonstrated drug-wise variation. Across studies, specificity (mean ≈ 0.92) exceeded sensitivity (mean ≈ 0.75). Bias diagnostics revealed marked funnel plot asymmetry; trim-and-fill imputation added 22, 29, and 23 studies for AUC, sensitivity, and specificity, respectively, yielding adjusted pooled estimates of ∼0.89 (AUC), 0.70 (sensitivity), and 0.93 (specificity).

CONCLUSIONS

ML models trained on genomic data demonstrate high diagnostic accuracy and robust discriminative ability for predicting first-line drug resistance-particularly for RIF and INH-although sensitivity remains variable across drugs and model types. Standardized external validation and calibration are needed before broad clinical deployment.