A risk calculator for multidrug-resistant tuberculosis in London, UK

Abstract

BackgroundRapid nucleic acid amplification technologies provide a fast and effective way to diagnose multidrug-resistant (MDR) tuberculosis. However, the number of false positive results produced by overtesting in low risk groups is a concern. With present estimates of the burden of MDR tuberculosis in the UK and rapid test sensitivity and specificity, the expected positive predictive value for testing all patients with tuberculosis is 0·47. Treatment for MDR tuberculosis is time-consuming and expensive, with severe side-effects for patients. Therefore an accessible risk calculator that can be used by clinicians to effectively target rapid testing is needed. MethodsUsing enhanced tuberculosis surveillance data collected through the London TB Register (a web-based system used by all TB clinics within the city) from 2009 to 2013, we developed a multivariable risk prediction model based on a-priori risk factors of MDR tuberculosis in London. Bootstrapping was used to internally validate the model, and external validation was done with national tuberculosis data from the same time period. FindingsOf 7292 individuals, 134 (1·83%) tested positive for MDR tuberculosis. In a multivariable penalised logistic regression model, the strongest risk factors associated with MDR tuberculosis included previous diagnosis, the burden of MDR tuberculosis in a patient's country of birth, recent arrival in the UK if foreign born, and age, with younger individuals at higher risk (appendix). The calibration slope was 0·90 after internal validation, and the area under the receiver operating characteristic curve (AUC) decreased from 0·73 to 0·70; this small reduction suggested satisfactory performance. Non-parametric calibration curves indicated accurate prediction for most patients, but with overestimated risk predicted for higher risk individuals. On an external dataset of 9652 patients, AUC was 0·74 with a calibration slope of 0·99. If only higher risk (>2%) patients were rapidly tested, we estimated that the positive predictive value would increase to 0·65. InterpretationThe model has been developed into a working calculator in Excel; the algorithms could also be integrated into surveillance registers to support embedding into clinical practice. Future research could include health economic work to help identify a range of cut-off risks at which testing is worthwhile. The model would need validating with local data before use in other countries. FundingOD is supported by a National Institute for Health Research (NIHR) Research Methods Fellowship.