Artificial Intelligence in Drug Susceptibility Testing for Multidrug-resistant Tuberculosis: Genomic Prediction, Automated Phenotyping and Clinical Translation

Abstract

Multidrug-resistant tuberculosis requires drug susceptibility information that is both rapid and sufficiently comprehensive to support effective regimen construction. Culture-based phenotypic drug susceptibility testing remains indispensable because it measures biological response directly, but its turnaround time, technical complexity and variable reproducibility for some drugs constrain timely treatment. Molecular testing is faster, yet fixed mutation catalogues and targeted assays can leave rare, complex or incompletely characterised resistance unresolved. Artificial intelligence, particularly machine learning and deep learning, is being developed as an interpretive layer for two complementary data streams: genomic sequence data and digital measurements of bacterial growth. This critical narrative review evaluates how these approaches may improve susceptibility prediction for multidrug- and rifampicin-resistant tuberculosis, where their evidence is strongest, and which methodological and implementation barriers limit clinical adoption. Genomic machine-learning models show the most mature evidence for rifampicin and isoniazid and increasingly for fluoroquinolones, while performance is less stable for drugs with sparse resistant training examples, uncertain genotype-phenotype relationships or difficult phenotypic reference standards. Quantitative models that predict minimum inhibitory concentrations may capture resistance gradations lost by binary classification, but their clinical interpretation remains immature. Computer-vision and single-cell growth approaches offer a different advantage: they retain a phenotypic readout while automating or accelerating culture interpretation. Their current evidence base, however, is smaller and includes important proof-of-concept studies that have not yet been validated in routine multidrug-resistant tuberculosis services. Across both domains, external validation, lineage and geographic representativeness, reference-standard quality, uncertainty reporting, reproducibility and workflow integration are more important determinants of clinical credibility than algorithmic complexity alone. Artificial intelligence should therefore be treated as an adjunct to validated laboratory methods, with explicit indeterminate outputs and reflex testing when predictions are uncertain. Prospective multicentre studies linking algorithm-assisted susceptibility testing to treatment decisions, turnaround time, patient outcomes, cost and equity are now the principal translational priority.