Data-driven approaches to hard-to-treat tuberculosis disease: a machine-learning based model for automated recommendation of individualized treatment

Background: In 2012, Bedaquiline, the first new antimicrobial for tuberculosis (TB) in 40 years’ time was approved. In 2021, the WHO endorsed bedaquiline, pretomanid, linezolid, and moxifloxacin (BPaLM) as the regimen of choice for the treatment of rifampicin resistant TB. As the global rollout of BPaLM continues, drug resistance will inevitably follow, and hard-to-treat TB will place a significant burden on health care systems, particularly in low resource settings with high incidence of drug-resistant TB (DR-TB). Most physicians have limited experience constructing effective individualized ‘rescue’ regimes for patients with hard-to-treat forms of TB characterized by resistance to BPaL drugs. We developed the DR-TB Treatment Recommender, an AI-based treatment decision aid that recommends the optimal DR-TB regimen based on the drug resistance profile of the infecting Mycobacterium tuberculosis strain. Methods: The development of the DR-TB treatment recommender system consisted of the assembly of the knowledge base, development of a heuristic model prototype, feedback from experts, application of machine learning methods to analyze the feedback, and assessment of the performance. For bedaquiline, we developed a Bayesian approach to determine the probability of bedaquiline resistance (pBDQR). We performed an online discrete choice experiment to investigate how pBDQR information influences the physician's decision to continue or stop BDQ. We also performed qualitative research to determine how physicians view the use of the treatment recommender. Results: pBDQR was the most influential factor affecting the BDQ prescribing decision of physicians, followed by response to treatment after 1 month. Based on these results, a system was developed to recommend continuation of BPaLM, strengthening the regimen, or individualizing the regimen, depending on the genomic drug resistance profile and the probability of resistance to BPaLM drugs. Following expert stakeholder meetings and literature review, nine drug features and 14 treatment regimen features were identified and quantified. Using machine learning, a first version for use prior to introduction of BPaLM was developed to predict the optimal treatment regimen based on a training set of 3895 treatment regimen-expert feedback pairs. A second version was developed for use in the context of BPaLM. Physician feedback highlighted the need for an improved approach for treatment recommendations for patients with bedaquiline, linezolid and/or pretomanid/delamanide resistance. Discussion: The Treatment Recommender is a valuable decision aid for healthcare workers treating patients suffering from DR-TB and could be critical in healthcare contexts with little or no access to expert committees. Further work is required to develop data-driven approaches capable of recommending optimal treatment regimes for patients with the hardest-to-treat TB, striking the right balance between efficacy and toxicity. Conclusion: Programs should consider incorporating this decision aid to faciliate the implementation of NGS-guided care and improve the treatment for patients with drug resistant TB.