Prediction Models for Tuberculosis Treatment Outcomes
Peetluk, Lauren Saag
Despite widespread availability of curative treatment, tuberculosis (TB) treatment outcomes remain sub-optimal, particularly among persons living with HIV. In 2018, global treatment success rates were 85% and 75% for TB and HIV-associated TB, respectively, far from the End TB Strategy target of ≥90%. To curtail TB morbidity and mortality, researchers and clinicians pursue early identification of individuals most likely to experience unsuccessful TB outcomes, as this may help direct interventions or resources towards those most in need. Clinical prediction modeling enables risk-based decision making and can lay the groundwork for future strategies to optimize TB outcomes. Additionally, given widespread recognition that HIV negatively influences TB outcomes, and the importance of isoniazid metabolism for safety and efficacy of TB therapy, these factors should be thoughtfully considered in prediction models. I systematically reviewed existing prediction models for TB treatment outcomes, including 33 studies presenting 37 models. Upon quality assessment, all models suffered bias, due to poor reporting of study population and data collection, complete-case analysis, univariate analysis-based model selection, lack of calibration assessment, and limited validation. Improving upon those results, I used data from 944 drug-susceptible, culture-confirmed, pulmonary TB cases enrolled in the Regional Prospective Observational Research for TB (RePORT) Brazil cohort to develop and internally validate a prediction model for unsuccessful TB treatment outcomes (death, treatment failure, loss to follow-up). The final model included seven baseline predictors: hemoglobin, HIV-infection, drug use, diabetes, age, education, and tobacco use. It demonstrated good discrimination (c-statistic=0·77; 95% confidence interval (CI): 0·73-0·80) and calibration (optimism-corrected intercept and slope: -0·12 and 0·89, respectively). I also estimated the incremental value of including HIV-related severity measures (CD4 T-cell count, HIV-1 RNA viral load, antiretroviral therapy use) or isoniazid acetylator status to provide insight about the importance of collecting these data, but none added notable improvement to the model. This study addresses the pressing need for improved prediction and prevention of unsuccessful TB outcomes. Methods adhered to expert guidance for developing prediction models. The resultant model performed well and can be easily implemented via a nomogram or web-based calculator.