Machine learning model helps determine resistance to key antibiotics for TB treatment

Tuberculosis (TB) is still among the top ten causes of death worldwide: in 2020, more than 1.3 million deaths were registered. The emergence and spread of drug-resistant forms of the disease have complicated the control of tuberculosis in many settings. An additional problem is the fact that the treatment of drug-resistant tuberculosis is complex (success rate was 57% in 2019), long-term (treatment can take 9-20 months) and multifaceted (treatment often requires several antibiotics that cause severe side effects).

The most important class of antibiotics for the treatment of drug-resistant tuberculosis are fluoroquinolones, which form the basis of most treatment-resistant tuberculosis regimens. However, tuberculosis strains have become resistant to fluoroquinolones, undermining the effectiveness of treatment regimens that include this class of antibiotics. The best treatment options for patients with drug-resistant tuberculosis are ideally determined by drug sensitivity tests that can phenotypically determine the effectiveness of antibiotics against a particular strain of tuberculosis. However, these tests are few in low-resource, high-load environments, which means that people in these regions cannot receive specialized treatment that can best treat their tuberculosis. Also, even if they are available, phenotypic testing can take up to 12 weeks to get results.

Reza Yaesubi, an associate professor of health at the Yale School of Public Health, and his team of researchers worked on models to predict fluoroquinolone resistance that could accelerate the process of providing optimal care. Working with national data on tuberculosis control collected in the Republic of Moldova, the team assessed whether demographic and clinical factors could be used as predictors of tuberculosis resistance to fluoroquinolones. They found that information such as age, geographic location, and whether the TB disease was new or recurrent served as reliable predictors of resistance. Based on this, they created a model using machine learning to assess the likelihood that a patient is infected with a fluoroquinolone-resistant strain of tuberculosis.

“One of their main advantages predictable patterns is that they can be deployed at the care point so clinicians can optimize treatment regimes, waiting for the results of drug sensitivity tests, which can take up to 12 weeks, ”Yaesubi said.

In contrast to the current strategy for the treatment of resistant tuberculosis, which initially assumes sensitivity to fluoroquinolones, the Yaesubi model considers how human circumstances affect the likelihood of resistance to fluoroquinolones and if alternative antibiotics (such as delamanide) should be used instead.

Through rigorous analysis and testing, the researchers found that the new model yields statistically higher net benefits when prescribing appropriate treatment for patients with drug-resistant tuberculosis. These findings promise a system that will better treat TB patients, Yasubi said. In the future, he hopes to extend the model beyond the data collected from the Republic of Moldova to include other regions with insufficient resources and high workloads.

“We plan to investigate whether similar prognostic models can be developed for other critical classes antibiotics and for other countries with high workloads drug resistant Tuberculosis, ”he said.

The study appears in PLOS Digital Health.