The effect of host factors on discriminatory performance of a transcriptomic signature of tuberculosis risk

Abstract

Background: We aimed to understand host factors that affect discriminatory performance of a transcriptomic signature of tuberculosis risk (RISK11).

Methods: HIV-negative adults aged 18-60 years were evaluated in a prospective study of RISK11 and surveilled for tuberculosis through 15 months. Generalised linear models and receiver-operating characteristic (ROC) regression were used to estimate effect of host factors on RISK11 score (%marginal effect) and on discriminatory performance for tuberculosis disease (area under the curve, AUC), respectively.

Findings: Among 2923 participants including 74 prevalent and 56 incident tuberculosis cases, percentage marginal effects on RISK11 score were increased among those with prevalent tuberculosis (+18·90%, 95%CI 12·66-25·13), night sweats (+14·65%, 95%CI 5·39-23·91), incident tuberculosis (+7·29%, 95%CI 1·46-13·11), flu-like symptoms (+5·13%, 95%CI 1·58-8·68), and smoking history (+2·41%, 95%CI 0·89-3·93) than those without; and reduced in males (-6·68%, 95%CI -8·31- -5·04) and with every unit increase in BMI (-0·13%, 95%CI -0·25- -0·01). Adjustment for host factors affecting controls did not change RISK11 discriminatory performance. Cough was associated with 72·55% higher RISK11 score in prevalent tuberculosis cases. Stratification by cough improved diagnostic performance from AUC = 0·74 (95%CI 0·67-0·82) overall, to 0·97 (95%CI 0·90-1·00, p < 0·001) in cough-positive participants. Combining host factors with RISK11 improved prognostic performance, compared to RISK11 alone, (AUC = 0·76, 95%CI 0·69-0·83 versus 0·56, 95%CI 0·46-0·68, p < 0·001) over a 15-month predictive horizon.

Interpretation: Several host factors affected RISK11 score, but only adjustment for cough affected diagnostic performance. Combining host factors with RISK11 should be considered to improve prognostic performance.

Funding: Bill and Melinda Gates Foundation, South African Medical Research Council.

Keywords: Host factors; Mycobacterium tuberculosis; Performance; RNA; Signature; Transcriptomic.

Figure 1

Predicted marginal effects on RISK11 by different host factors in (a) all participants, (b) participants with prevalent tuberculosis, and (c) participants without tuberculosis. Prev TB, Prevalent tuberculosis. N Sweat, Night sweats. Smoking, Smoking history. Flu-like, Flu-like symptoms. BMI, Body-mass index. Incid TB, Incident tuberculosis. Prior TB, Prior tuberculosis. The midline indicates the percentage marginal effect and the error bars indicate the 95% CIs.

Figure 2

Crude and covariate-adjusted ROC curves for the discrimination of (a) prevalent TB from controls and (b) incident TB from controls. The crude and covariate-adjusted ROC curves are superimposed in both figures (a) and (b). The ROC curves are adjusted for BMI, sex, night sweats, haemoptysis, flu-like symptoms, and smoking history in both instances.

Figure 3

Performance of RISK11 when stratified by cough status and when combined with host factors. (a) Covariate-specific ROC curves for the diagnostic performance of RISK11 in cough-positive n = 58) and cough-negative (n = 2,865) individuals. Crude (RISK11 only), and combination ROC curves for discriminating (b) prevalent TB versus controls, and (c) incident TB versus controls through 15 months follow-up. Baseline model AUCs were derived from predictive models containing age, BMI, and cough for prevalent TB; and BMI smoking history, and previous TB history for incident TB.