Machine Learning for Predicting Drug Resistance in Tuberculosis: A Systematic Review of Model Performance and Clinical Applications

This systematic review evaluates the performance of diverse Machine Learning (ML) models in predicting drug resistance in Mycobacterium tuberculosis (M. tuberculosis), covering both first-line and second-line anti-Tuberculosis (anti-TB) drugs. The study analyzes models, such as Hierarchical Attentive Neural Network (HANN) variants, MultiDrug-Wide Deep Neural Network (MD-WDNN), Deep learning for AntiMicrobial Resistance (DeepAMR), Logistic Regression (LR), Random Forest (RF), and TuBerculosis Drug Resistance Optimal Prediction (TB-DROP), assessing their accuracy, sensitivity, specificity, and Area Under the Curve (AUC) across multiple datasets. Top-performing models demonstrate high discriminative power for key first-line drugs. Notably, HANN with variant encoder as Transformer and gene encoder as Transformer (HANN-TT) achieves the highest sensitivity for rifampicin (96.31%), and HANN-BB and Treesist-TB excelling in predicting isoniazid resistance, with specificity values reaching 99.2%. Classification Tree using ALL SNPs (CT-ALL) and Gradient Boosted Tree using Co-occurrence Resistance Markers (GBT-CRM) also produce strong results for ciprofloxacin and pyrazinamide, respectively. These findings underscore the adaptability of ML models to varied drug resistance patterns and their clinical value in early detection and treatment planning. However, considerable variability is observed in predicting resistance to ethambutol and pyrazinamide, reflecting challenges due to complex mutation profiles and limited annotated data. Overall, ML models demonstrate strong potential to transform TB diagnostics through rapid, accurate resistance profiling. For effective clinical integration, large-scale validation, enhanced training datasets, and context-specific deployment strategies are required. This review offers critical insights to inform future model development, with an emphasis on integrating genomic, clinical, and demographic data for more personalized and effective TB care.