Nationwide longitudinal evaluation of a machine learning approach for enhanced interpretation of Xpert MTB/RIF ultra rifampicin-resistance results in low bacterial load tuberculosis specimens.

BACKGROUND

The World Health Organization (WHO) has identified tuberculosis (TB) as the leading cause of death from a single infectious agent. False-positive rifampicin (RIF) resistance results from the Xpert MTB/RIF Ultra assay are common in TB patients with low bacterial loads, especially among HIV-coinfected individuals. Hence, to distinguish genuine RIF resistance from false-positive results, this study developed and validated an artificial intelligence clinical decision support system (AI-CDSS).

METHODS

Between January 2021 and March 2025, Taiwan's national TB reference laboratory received 10,353 respiratory specimens nationwide, identifying 2443 MTB-positive samples. The specimens were subjected to Xpert MTB/RIF Ultra testing and RIF resistance was confirmed using GenoType MTBDRplus assays. Molecular features, including cycle threshold (Ct) values, melting temperatures (Tm), and fluorescence intensities of rpoB probes, were analyzed. Three machine learning algorithms: random forest, gradient boosting classifier, and light gradient boosting machine (LGBM) were trained and validated.

RESULTS

Ultra initially reported RIF resistance in 174 samples (7.1 %), with the highest false-positive rate of 12.2 % observed in samples with very low bacterial loads. LGBM demonstrated superior diagnostic performance (AUC = 0.99, sensitivity = 0.97, specificity = 0.99, and F1-score = 0.98). Key predictive features included Tm and fluorescence intensity, particularly in the rpoB3 region. Implementing the AI-CDSS significantly improved accuracy and reduced diagnostic turnaround times.

CONCLUSIONS

By leveraging the LGBM model, AI-CDSS effectively distinguished true RIF resistance from false-positive Xpert Ultra results, particularly among patients with low MTB bacterial loads. This approach enhances clinical decision making, optimizes treatment initiation, and conserves vital multidrug-resistant TB resources.