Using Machine Learning Methods to Predict Early Treatment Outcomes for Multidrug-Resistant or Rifampicin-Resistant Tuberculosis to Enhance Patient Cure Rates: Development and Validation of Multiple Models.

Background: Early prediction of treatment outcomes for patients with multidrug-resistant or rifampicin-resistant tuberculosis (MDR/RR-TB) undergoing extended therapy is crucial for enhancing clinical prognoses and preventing the transmission of this deadly disease. However, the absence of validated predictive models remains a significant challenge.

Objective: This study compared a conventional logistic regression model with machine learning (ML) models using demographic and clinical data to predict outcomes at 2 and 6 months of treatment for MDR/RR-TB. The goal was to advance model applications, refine control strategies, and boost MDR/RR-TB cure rates.

Methods: This retrospective study encompassed an internal cohort of 744 patients with MDR/RR-TB examined between January 2017 and June 2023, as well as an external cohort comprising 137 patients with MDR/RR-TB examined between March 2021 and June 2022. Data on culture conversion were collected at 2 and 6 months, and culture conversion was tracked in the external cohort at the same time points. The internal cohort was assigned as the training set, whereas the external cohort was used as the validation set. Logistic regression and 7 ML models were developed to predict the culture conversion of patients with MDR/RR-TB at 2 and 6 months of treatment. Model performance was evaluated using the area under the curve, accuracy, sensitivity, and specificity.

Results: In the internal cohort, culture conversion rates for MDR/RR-TB were 81.9% (485/592) at 2 months and 87.1% (406/466) at 6 months. The odds ratio for treatment success was 8.55 (95% CI 3.31-22.08) at 2 months and 20.33 (95% CI 6.90-59.86) at 6 months after conversion, with sensitivities of 86.5% and 92.2% and specificities of 57.1% and 63.2%, respectively. The artificial neural network model was the best for culture conversion at both 2 and 6 months of treatment, with areas under the curve of 0.82 (95% CI 0.77-0.86) and 0.90 (95% CI 0.86-0.93), respectively. The accuracy, sensitivity, and specificity of the model were 0.74, 0.74, and 0.75 at 2 months of treatment and 0.80, 0.79, and 0.87 at 6 months of treatment, respectively.

Conclusions: The ML models based on 2- and 6-month culture conversion could accurately predict treatment outcomes for patients with MDR/RR-TB. ML models, particularly the artificial neural network model, outperformed the logistic regression model in both stability and generalizability and offer a rapid and effective tool for evaluating therapeutic efficacy in the early stages of MDR/RR-TB treatment.