Systematic review and meta-analysis of predictive accuracy of prognostic models for poor treatment outcome of drug resistance tuberculosis.

Tuberculosis (TB) remains a global health challenge, with drug-resistant (DR) forms of the disease posing a significant threat to effective disease control. Prognostic models are crucial tools that aid in clinical decision-making by predicting the likelihood of adverse treatment outcomes. Evaluating the accuracy of these models is essential to ascertain their reliability and effectiveness in guiding healthcare interventions and optimizing patient outcomes. This study aimed to systematically review and meta-analyze the predictive accuracy of the DR-TB poor treatment outcome prediction models. A systematic search was conducted in four databases (Scopus, Embase, PuBMed, and HINARI) to identify studies based on Population, Index model, Comparator, Outcome, Timing, and Setting (PICOTS) approach until May 20, 2023. We extracted data using the CHARMS checklist and appraised risk of bias using PROBAST tool. Discrimination and calibration performance were meta-analyzed when appropriate. A total of 11 studies with 13 models were found to be eligible for this study. The cohort sizes vary from 102 to 2,441 participants involved in model development, while the number of events per parameter (EPP) in the model remains consistently below 10 across the majority of studies. The individual studies underwent assessment for Risk of Bias (ROB) and Applicability using PROBAST. The primary reason for the high concern of ROB was predominantly within the Analysis domain. The pooled Area Under the Curve (AUC) from the Receiver Operating Characteristic (ROC) analysis of the eligible models was 0.77 (95% CI 0.73, 0.81), which is within an acceptable range. The poor treatment outcomes of DR-TB prediction models manifest in various forms, characterized by diverse predictors and methodological approaches. While the pooled predictive accuracy of eligible models is deemed acceptable, concerns arise regarding a higher risk of bias, notably within the analysis domain, and a deficiency in external validation. Authors of such prediction models are encouraged to enhance their models by incorporating external validation and conducting clinical utility assessments.