Tefera Agizew, Yasmeen Hanifa, Heather Alexander, Unami Mathebula-Modongo, Yuliang Liu, Anand Date, Rosanna Boyd, Andrew D. Kerkhoff, Andrew F. Auld, Robin Wood, George Bicego, Anikie Mathoma, Goabaone Rankgoane-Pono, Alison D. Grant, Salome Charalambous, Alyssa Finlay, Pontsho Pono, Katherine Fielding, Tedd V. Ellerbrock, James Shepherd, Ray W. Shiraishi, and Christopher Serumola
Background Among people living with HIV (PLHIV), more flexible and sensitive tuberculosis (TB) screening tools capable of detecting both symptomatic and subclinical active TB are needed to (1) reduce morbidity and mortality from undiagnosed TB; (2) facilitate scale-up of tuberculosis preventive therapy (TPT) while reducing inappropriate prescription of TPT to PLHIV with subclinical active TB; and (3) allow for differentiated HIV–TB care. Methods and findings We used Botswana XPRES trial data for adult HIV clinic enrollees collected during 2012 to 2015 to develop a parsimonious multivariable prognostic model for active prevalent TB using both logistic regression and random forest machine learning approaches. A clinical score was derived by rescaling final model coefficients. The clinical score was developed using southern Botswana XPRES data and its accuracy validated internally, using northern Botswana data, and externally using 3 diverse cohorts of antiretroviral therapy (ART)-naive and ART-experienced PLHIV enrolled in XPHACTOR, TB Fast Track (TBFT), and Gugulethu studies from South Africa (SA). Predictive accuracy of the clinical score was compared with the World Health Organization (WHO) 4-symptom TB screen. Among 5,418 XPRES enrollees, 2,771 were included in the derivation dataset; 67% were female, median age was 34 years, median CD4 was 240 cells/μL, 189 (7%) had undiagnosed prevalent TB, and characteristics were similar between internal derivation and validation datasets. Among XPHACTOR, TBFT, and Gugulethu cohorts, median CD4 was 400, 73, and 167 cells/μL, and prevalence of TB was 5%, 10%, and 18%, respectively. Factors predictive of TB in the derivation dataset and selected for the clinical score included male sex (1 point), ≥1 WHO TB symptom (7 points), smoking history (1 point), temperature >37.5°C (6 points), body mass index (BMI) 10) yielded TB prevalence of 1%, 1%, 2%, and 6% in the lowest risk group and 33%, 22%, 26%, and 32% in the highest risk group for XPRES, XPHACTOR, TBFT, and Gugulethu cohorts, respectively. At clinical score ≥2, the number needed to screen (NNS) ranged from 5.0 in Gugulethu to 11.0 in XPHACTOR. Limitations include that the risk score has not been validated in resource-rich settings and needs further evaluation and validation in contemporary cohorts in Africa and other resource-constrained settings. Conclusions The simple and feasible clinical score allowed for prioritization of sensitivity and NPV, which could facilitate reductions in mortality from undiagnosed TB and safer administration of TPT during proposed global scale-up efforts. Differentiation of risk by clinical score cutoff allows flexibility in designing differentiated HIV–TB care to maximize impact of available resources., Andrew Auld and colleagues evaluate a clinical score for active tuberculosis in persons with HIV infection., Author summary Why was this study done? Tuberculosis (TB) remains the most common cause of death among people living with HIV (PLHIV) and is often undiagnosed at time of death. Rapid scale-up of tuberculosis preventive therapy (TPT) to 13 million PLHIV in low- and middle-income countries (LMICs) has been proposed for 2021; however, active TB is commonly asymptomatic and therefore missed by current WHO-recommended 4-symptom TB screening rules. Therefore, more sensitive TB screening tools are needed to better facilitate early TB diagnosis and safer scale-up of TPT to PLHIV by avoiding TPT prescription to clients with asymptomatic active TB, who need TB treatment. What did the researchers do and find? We derived a TB risk score for PLHIV from XPRES trial data and validated the score on 3 external datasets. We prioritized high sensitivity and ability to correctly rule out TB (i.e., high negative predictive value (NPV)) at key time points in care such as HIV clinic enrollment and before TPT prescription. Both logistic regression and random forest machine learning approaches were used to identify the 6 most important predictors, commonly available in LMIC clinic settings. In the external datasets, TB risk score ≥2 had higher sensitivity (87% to 97%) than WHO 4-symptom screening rule and increased NPV by 0.3% to 1.7%. Three risk groups were identified by the score, with active TB prevalence in external datasets ranging from 1% to 6% in the lowest to 22% to 32% in the highest risk groups. What do these findings mean? Following further validation, this clinical score could improve early detection of active TB to reduce morbidity and mortality from undiagnosed TB. Use of the clinical score cutoff of ≥2 during the proposed TPT scale-up for 13 million PLHIV could potentially avoid many thousands of PLHIV with active TB being inappropriately prescribed TPT. By differentiating 3 risk groups, the score also allows for the development of differentiated service delivery models suitable for LMIC.