Investigators provided data about 29,523 participants, of whom 10,057 were people living with HIV. The dataset included 9,626 people living with HIV who had TB screening and sputum culture performed, of whom 8,148 could be evaluated on the five symptoms of interest from nine of 12 studies ().
Flow chart of study participants included in the individual patient data meta-analysis.
Most patients (77% [7,386/9,626]) were from sub-Saharan Africa; the rest were from Southeast Asian countries. The median age was 34 y (interquartile range [IQR], 27–41 y). Of the 9,626 patients with HIV in the 12 studies, CD4 cell count information was available for 3,489 (36%) and chest radiography information for 3,903 (41%). The median CD4 count was 248 cells/µl (IQR, 107–409).
The overall prevalence of TB disease was 5.8% (557/9,626), ranging across studies from 0.4% to 25.7% (). More than half of TB patients (52% [288/557]) had sputum smear negative pulmonary TB, whereas 39% (218/557) had sputum smear positive pulmonary, and 5% (28/557) had exclusively extrapulmonary TB. The anatomic site of TB was not specified in 4% (23/557) of patients.
summarizes the distribution of common variables, and Table S1
summarizes how each question was actually asked in each study. Because duration of cough was included in many studies but was asked about in different ways, we were able to analyze data using three different cough variables: cough in the past 4 wk (information available for 39.3% of participants); cough lasting for 2 wk or more (information available for 47.1%); and cough present in the last 24 h, which is referred to as “current cough” (information available for 89.6%).
Characteristics of participants with and without TB for variables included in the analysis.
We analyzed the performance of individual and combinations of symptoms as screening rules using data from the 8,148 participants who could be evaluated based on the five candidate symptoms. shows the diagnostic performance characteristics for the 23 candidate combinations of symptoms, sorted from highest sensitivity to lowest. The most sensitive rule was the presence of any one of the following symptoms: current cough, fever, night sweats, and weight loss (CFSW). The population-average sensitivity of this symptom combination was 78.9% (95% confidence interval [CI] 58.3%–90.9%) with the negative likelihood ratio of 0.426 (95% CI 0.349–0.520), which corresponds to a postscreening reduction in the probability of TB by 15%–20% 
The nine included studies demonstrated significant between-study heterogeneity on both sensitivity (p
<0.001) and specificity (p
<0.001) of the rule CFSW (see also ). The bivariate graphic shows that six studies have study-level specificities below and three above the population average specificity. Furthermore, this rule has the highest-ranking sensitivity in eight of the nine included studies (Table S2
). The hierarchical summary relative operating characteristic curves (Figure S1
) show slightly better overall diagnostic performance of the rules CFS and CF, but our application requires the highest sensitivity possible, allowing for some tradeoffs with lower specificity. shows that three studies are outliers, and they represent studies of patients who were previously screened for TB or studies in which much of the population likely had previous TB screening (e.g., miners); this can modify the performance characteristics of the screening rule.
Diagnostic performance of CFSW rule in the included studies.
The CFSW rule has sensitivity of 90.1% (95% CI 76.3%–96.2%) and 67.1% (95% CI 41.7%–85.3%) among participants selected from clinical and community settings, respectively. Similarly the sensitivity of the rule among those who had not been previously screened for TB was higher at 88.0% (95% CI 76.1%–94.4%) compared to those who had been screened for TB at 40.5% (95% CI 16.6%–69.9%). At the 95% confidence level, the sensitivity of this rule could not be statistically distinguished from the sensitivity of the rule that substitutes haemoptysis for current cough (HFSW, 75.7% sensitive [95% CI 53.9–89.2%]) or the rule that drops night sweats (CFW, 74.0% sensitive [51.7–88.3%]). All other rules had lower sensitivity.
Regression analysis of study-level predictors revealed that studies in which TB screening was performed in clinical settings had 4.5 times the odds for a true-positive screening result compared to studies in which TB screening was performed in a community setting (95% CI 1.0–19.5). Studies of participants who had not previously been screened for TB had 10.8 times the odds for a true-positive screen (95% CI 2.4–47.8) compared with studies in which participants had previously been screened for TB. Participants with CD4 cell count <200 cells/µl had 6.4 times the odds of a true-positive screen (95% CI 2.9–14.2). Statistically significant predictors of true-negative results include prescreening, geographic region, participant age ≥33 y, CD4 cell count <200 cells/ml, and abnormal result on chest radiograph ().
Association of study-level and individual-level predictors with the diagnostic performance of CFSW rule.
shows the negative predictive value and the numbers needed to screen for the CFSW rule adjusted for individual- and study-level covariates. In a setting with 5% TB prevalence among people living with HIV, the rule has a negative predictive value of 98.3% (95% CI 97.5%–98.8%) for patients screened in a clinical setting and 97.3% (95% CI 96.9%–97.7%) for patients screened in a community setting. The numbers needed to screen at the same prevalence of TB are 15 and 11 for clinical and community setting, respectively. The negative predictive value was similar in those having high (≥200) and low (<200) CD4 count at 96.9% (95% CI 95.1%–98.0%) and 98.9% (95% CI 97.5%–99.5%), respectively (see also Table S3
Negative predictive value (NPV) and number needed to screen (NNS) using rule CFSW in a hypothetical population of 1,000 people living with HIV stratified by study and individual level predictors.
Four studies 
consistently recorded information on chest radiograph, allowing screening rules with this sign to be evaluated using data from 2,805 participants The addition of abnormal chest radiographic findings into the CFSW rule increases the sensitivity to 90.6% (95% CI 66.7%–97.9%) with a specificity of 38.9% (95% CI 12.8%–73.3%), and a likelihood ratio negative of 0.242 (95% CI 0.102–0.571). Fifteen of the 23 rules included in our analysis outperform the symptom-based CFSW rule when abnormal chest radiographic findings are added (Table S4
On the basis of our meta-analysis findings and incorporating current WHO recommendations on provision of IPT, we developed a simple TB screening algorithm for public health programmes to screen people living with HIV, and, depending on the outcome of screening, to either provide IPT or evaluate patients further for TB or other diseases ().
Algorithm for TB screening in person living with HIV in HIV prevalent and resource-constrained settings.