A Treatment-Decision Score for HIV-Infected Children With Suspected Tuberculosis

The ANRS 12229 PAANTHER 01 Study was a cohort study aimed at developing an algorithm to improve diagnosis of tuberculosis in children infected with HIV that was conducted in 8 hospitals in Burkina Faso, Cambodia, Cameroon, and Vietnam (April 2011–December 2014) (Supplemental Methods section of the Supplemental Information). Inclusion procedures and the study design have been described elsewhere.²⁵  In brief, we enrolled HIV-infected children aged ≤13 years with suspected tuberculosis on the basis of at least 1 of the following: (1) persistent cough; (2) fever for >2 weeks; (3) failure to thrive, defined as recent deviation in the growth curve or a weight-for-age z score (WAZ) <−2 SDs; (4) failure of antibiotics for a pulmonary infection; or (5) a suggestive chest radiograph (CXR). We excluded those with antituberculosis treatment started within 2 years before inclusion.

The study was approved by relevant national ethics committees, institutional review boards, and national authorities. The ANRS 12229 PAANTHER 01 Study is registered at ClinicalTrials.gov (identifier NCT01331811).

After parent or guardian informed consent, a detailed history on presence and duration of symptoms 4 weeks before enrollment was collected from parent(s) or guardian(s) through a standardized questionnaire. Cough patterns were characterized by using a graphic illustration shown to parent(s) or guardian(s).²⁶  All children had a complete physical examination; CXR, abdominal ultrasonography, and TST performed; and blood samples collected for HIV RNA, CD4, complete blood cell count, transaminases, and QFT. Each child had 2 to 3 gastric aspirates or expectorated sputa, 1 nasopharyngeal aspirate, 1 stool sample, and 1 string test, if aged ≥4 years, collected over a period of 3 days for Xpert, smear microscopy, and a mycobacterial culture.²⁵  ART and antituberculosis treatment were initiated at the discretion of the treating physician. All children were followed-up for 6 months. All data were collected by using standardized paper case-report forms and entered in an online database developed on the Voozanoo software (Epiconcept, Paris, France).

At the end of the study, children were retrospectively classified as having confirmed, unconfirmed, or unlikely tuberculosis by using the updated Clinical Case Definition for Classification of Intrathoracic Tuberculosis (Supplemental Table 4).²⁷  For model development, the reference diagnosis was tuberculosis, defined either as confirmed or unconfirmed. CXRs were reviewed independently by 2 readers blinded to patient data; discordant opinions were resolved by a third reader. Results of the TST were considered positive if the transverse diameter of induration, read at 48 to 72 hours, was >5 mm. QFT results, interpreted per the manufacturer’s recommendation, were not taken into account for the reference diagnosis. We used age-defined standards for tachycardia and tachypnea (Supplemental Table 5).²⁸  Sample-size calculations are detailed in the Supplemental Methods section of the Supplemental Information.

We compared baseline characteristics between groups using Student’s t test or the Kruskal-Wallis test and Pearson’s χ² or Fisher’s exact test, as appropriate.

We used logistic regression to develop diagnostic prediction models for tuberculosis. We restricted the analysis to those children with data available for candidate predictors. We included, as candidate predictors, characteristics used in previous childhood tuberculosis scoring systems and characteristics previously described as associated with tuberculosis in children infected with HIV as well as QFT and abdominal ultrasonography results (Supplemental Methods section of the Supplemental Information).^{22,23,29–34}  To identify additional predictors, we performed a nested case-control study, selecting as case patients all children with culture-confirmed tuberculosis and as controls those with unlikely tuberculosis who were still alive at month 6 and had not been treated for tuberculosis. We included as predictors CXR features, as assessed by the local reader. We tested various symptom durations (>2, >3, and >4 weeks) in the models and selected the one with the best Akaike information criterion. We also included ART and immunodeficiency as predictors in the models and tested interactions with other predictors.

To account for the fact that QFT and abdominal ultrasonography may not be systematically available in low- and middle-income countries (because they were not recommended by the WHO for tuberculosis diagnosis), we developed 4 different models: (1) all predictors integrated, (2) QFT excluded, (3) abdominal ultrasonography excluded, and (4) both QFT and abdominal ultrasonography excluded. We obtained final models by stepwise backward selection. As recommended for prediction models, we used less stringent P values of <.157 or .135 when incorporating variables with 1 or 2 degrees of freedom to avoid overfitting and to reduce model optimism.³⁵  We included Xpert and smear microscopy results secondarily in final models using Firth’s penalized likelihood to solve the problem of data separation.³⁶  We performed internal validation using bootstrap resampling (Supplemental Methods section of the Supplemental Information).³⁵ 

We compared areas under the receiver operating characteristic curves (AUROCs) of models obtained and selected the best model on the basis of discriminative ability and parsimony. We developed an associated diagnostic score by assigning to each variable category a predictor score equal to its β coefficient in the model. We set the tuberculosis diagnosis threshold using a predicted probability cutoff that reached a sensitivity of 90% in the case-control subpopulation. To facilitate final score calculations, we multiplied all predictor scores by a factor setting the threshold to 100 (Supplemental Table 12).

We assessed diagnostic performance of the score obtained in the whole cohort, considering missing data for predictors as all negative or all positive. Finally, we proposed a diagnostic algorithm that included the score.

We performed all analyses using SAS software version 9.3 (SAS Institute, Inc, Cary, NC).

We enrolled 438 children in the study (Table 1). Tuberculosis was confirmed by culture and/or Xpert in 55 (12.6%) children, and 196 (44.7%) children were classified as having unconfirmed tuberculosis, for a total of 251 (57.3%) children with tuberculosis.

Individual predictors with the best sensitivity for tuberculosis diagnosis were cough in the previous 4 weeks, cough lasting >2 weeks, cough in the past 24 hours, fever in the previous 4 weeks, and weight loss, as reported by the parent(s) or guardian(s), in the previous 4 weeks (Table 2). Specificities of these signs were poor overall.

A total of 335 of 438 children had data available for all selected predictors and were included in model development, including 201 (60.0%) children with tuberculosis and 134 (40.0%) who were classified as not having tuberculosis. Compared with children who were not included in model development, children with all predictors available were older, had a higher WAZ and higher hemoglobin count, more frequently had nontuberculous mycobacteria isolated, and had antituberculosis treatment initiated, and their risk of death was lower (Supplemental Tables 6 and 7).

Of predictors associated with tuberculosis diagnosis in the case-control subanalysis, which included 45 culture-confirmed tuberculosis cases and 153 control cases (Supplemental Results section of the Supplemental Information, Supplemental Tables 8 and 9), tachycardia only was not part of our initial list of predictors considered (Supplemental Table 10).

An identical set of 9 predictors remained in the 4 models: fever lasting >2 weeks, unremitting cough, hemoptysis and weight loss in the previous 4 weeks, contact with a patient with smear-positive tuberculosis, tachycardia, miliary tuberculosis, alveolar opacities, and lymph nodes on the CXR (Supplemental Table 11). ART and immunodeficiency did not improve model predictions and thus were not included in the final model; QFT results and abdominal lymph nodes on the ultrasound were ultimately included in the final models because they improved model predictions significantly. There was no interaction between ART or immunodeficiency and other predictors.

AUROCs for models 1, 2, 3, and 4 were 0.839 (95% confidence interval [CI] 0.797–0.880), 0.830 (95% CI 0.787–0.873), 0.819 (95% CI 0.775–0.863), and 0.808 (95% CI 0.762–0.853), respectively. Compared with model 1, only model 4 had a significantly lower discriminative ability (P = .220, P = .072, and P = .0191 for models 2, 3, and 4, respectively).

Including smear microscopy as a predictor did significantly change the discriminative ability of models 1, 2, and 3, compared with models without smear microscopy. It only led to a significant AUROC increase for model 4 (Supplemental Results section of the Supplemental Information). Including Xpert in the final models led to better discriminative ability for all models, with significant increases in AUROCs to 0.866 (95% CI 0.829–0.904), 0.861 (95% CI 0.822–0.899), 0.850 (95% CI 0.809–0.890), and 0.846 (95% CI 0.805–0.887) for models 1, 2, 3, and 4, respectively, compared with models without Xpert (P = .0015, P = .0011, P = .0007, and P = .0002) (Fig 1), without changes in other model predictors selected (Table 3). Compared with model 1, only model 4 had a significantly lower discriminative ability (P = .2800, P = .0799, and P = .0499 for models 2, 3, and 4, respectively). Model optimism was estimated to 0.0584 (95% CI 0.0147–0.0872) for model 2, which had the best discriminative ability and parsimony.

The score was developed on model 2 (Supplemental Table 13) by using the predicted probability cutoff that obtained a sensitivity >90% in the case-control population (Supplemental Figs 3 and 4, Supplemental Table 12). It had the following diagnostic accuracy measures: sensitivity: 178 of 201 (88.6%; 95% CI 84.2%–93.0%); specificity: 82 of 134 (61.2%; 95% CI 52.9%–69.4%); positive predictive value: 77.4% (95% CI 72.0%–82.8%); negative predictive value: 78.1% (95% CI 70.2%–86.0%).

The score sensitivity did not differ between the 4 countries (P = .144); specificities were significantly lower in Cambodia and Cameroon (43.3% [95% CI 25.6%–61.1%]; 40% [95% CI 23.8%–56.2%]; P < .0001) (Supplemental Results section of the Supplemental Information). Sensitivity did not differ between patients with a CD4 percentage <10% and those with a CD4 percentage ≥10% (P = .568); specificity was lower in those with a CD4 percentage <10% (53.1%; 95% CI 39.1%–67.0%; P = .014). Sensitivity and specificity did not differ between children with chronic cough as an inclusion criterion and the others (P = .838 and P = .485).

The score applied to the overall cohort correctly identified 228 (85.7%) children with tuberculosis and 116 (62.0%) children without tuberculosis when all missing predictors were considered as negative. Conversely, when all missing data were considered as positive, it correctly identified 228 (90.8%) children with tuberculosis and 82 (43.9%) children without tuberculosis.

In children infected with HIV presenting with a clinical suspicion of tuberculosis based either on chronic cough for >2 weeks or other study eligibility criteria, including a suggestive CXR (if done previously), the score can be applied in a stepwise approach (Fig 2). Antituberculosis treatment should be initiated immediately in children with a score of >100. Tuberculosis may be ruled out in children who score below 100 after full assessment, with a recommended subsequent clinical reassessment for persistent symptoms. If abdominal ultrasonography cannot be available, an alternative score may be used (Supplemental Table 14), with a sensitivity of 90.0% (95% CI 85.9%–94.2%) but a lower specificity of 48.5% (95% CI 40.0%–57.0%).

We developed a diagnostic prediction score for antituberculosis treatment decision in HIV-infected children with suspected tuberculosis. We aimed to provide clinicians from high–tuberculosis burden and resource-limited settings with a decision-making tool to initiate antituberculosis treatment quickly in HIV-infected children with suspected tuberculosis. The score obtained had a sensitivity of ∼90% and a specificity of 61%.

To our knowledge, this is the first study in which a diagnostic score is developed exclusively in children infected with HIV by using methods recommended for diagnostic prediction models. Previous pediatric tuberculosis diagnostic scores and algorithms were mostly based on expert opinion and often lacked validation.^13,14  A prospective study in South Africa revealed that the combined presence of 3 symptoms constituted a good tuberculosis diagnostic approach in HIV-uninfected children aged ≥3 years, but it performed poorly in those infected by HIV (sensitivity: 56%; specificity: 62%).²²  Recently, a retrospective study of scoring systems in Brazilian children infected with HIV who were evaluated for tuberculosis revealed that an extended version of the South African approach had a sensitivity of 94% or 84%, depending on whether a microbiologic evaluation was included in the evaluation, and a specificity of 30%.¹⁸  Our score therefore has good performances overall, compared with these scoring systems, with a good sensitivity and an acceptable specificity if Xpert is used. Despite increasing availability of the GeneXpert platform in high–tuberculosis burden countries, access to Xpert may still be challenging in some resource-limited settings.³⁷ 

The vast majority of children had prolonged cough as an inclusion criterion, which therefore lacked specificity. However, unremitting cough, which was assessed by using a graphic illustration, revealed a much higher specificity and remained in the model. Tachycardia, which was not used before in pediatric tuberculosis scores, is part of the 3 danger signs that should trigger antituberculosis treatment initiation in severely ill adults infected with HIV.³⁸  CXR findings significantly contributed to our score; yet, there is limited access to quality CXR and lack of reading skills in limited-resource settings. We used the local reader’s opinion, which constituted an imperfect but more practical test compared with more experienced readers.³⁹  Presence of lymph nodes on the ultrasound had similar diagnostic accuracy compared with that found in South African children infected with HIV and significantly improved the model’s discriminative ability.^23,24  Sensitivity of QFT in our study was much lower than the pooled sensitivity estimated at 47% in a recent meta-analysis in children infected with HIV and did not improve the model’s discriminative ability, confirming its poor diagnostic performance for tuberculosis in children with immunodeficiency.⁹ 

With its high sensitivity, our score should enable standardized treatment initiation in most HIV-infected children with tuberculosis. We showed previously that mortality in ART-naïve children was associated with the lack of treatment rather than the delay to antituberculosis treatment.¹²  However, initiation of antituberculosis treatment within a median of 1 week led to delayed ART, which was associated with increased mortality. It was recently estimated that in high–tuberculosis burden countries, it may be more cost-effective to treat all children with presumptive tuberculosis.⁴⁰  In children infected with HIV, however, pill burden and potential impact on ART have led to a call for a more discriminant approach. With the step-by-step approach, the score could enable same-day treatment decision without CXR and abdominal ultrasonography in children presenting clinical criteria. Overall, our score did not perform as well as clinicians from study tertiary health care facilities who treated 92% of children with tuberculosis and only 6% of those without; however, we expect that it will contribute to faster treatment decision at lower levels of care, especially when used with feasible and sensitive specimens for Xpert, such as nasopharyngeal aspirates and stools.²⁵  In practice, access to treatment does not depend exclusively on treatment decision and may be delayed for other structural reasons.

Our study has limitations. First, an incorporation bias resulting from the lack of a reference standard for childhood tuberculosis, independent from candidate predictors, may have led to overestimation of the models’ diagnostic performance.⁴¹  The good discriminative ability of the model in the case-control subset, however, reveals limited impact on the score performances. Second, almost one-quarter of study participants, mostly younger children with severe clinical status, had missing data for the considered predictors. Our analysis reveals, however, that the score has similar sensitivity in these children and that missing data would mostly impact specificity, which varied between 43% and 61%. Lastly, our study eligibility criteria differed from WHO criteria for investigation of tuberculosis, namely poor weight gain, fever, current cough, and history of contact with a patient with tuberculosis.⁴²  Our score is therefore not directly applicable to children presenting with these criteria. Despite these limitations our study has strengths. Development by using data from 4 countries ensured better external validity and generalizability of the scores, and internal validation revealed that the models developed would provide good predictions.⁴³  The lower score specificity in Cambodia could be due to higher rates of nontuberculous mycobacteria disease, which is difficult to distinguish from tuberculosis.⁴⁴ 

With its high sensitivity and algorithmic approach, the PAANTHER score should enable rapid treatment decision in children with presumptive tuberculosis. This algorithm constitutes a consequentialist approach to tuberculosis in children infected with HIV, considering the need to initiate treatment to reduce mortality, rather than an essentialist approach, considering the trueness of tuberculosis diagnosis.⁴⁵  However, further external validation is needed to validate both the scoring system and the overall approach and to confirm its clinical usefulness.

We thank all children and their parents and caregivers for their participation in the study, national tuberculosis and HIV programs from participating countries for their support, Françoise Barré-Sinoussi and Jean-François Delfraissy for their continuous support, Xavier Anglaret for general guidance, Julien Asselineau and Paul Perez for methodologic advice, and Corine Chazallon and Vincent Bouteloup for statistical support.

ART

antiretroviral therapy

AUROC

area under the receiver operating characteristic curve

CI

confidence interval

CXR

chest radiograph/radiography

QFT

Quantiferon Gold In-Tube

TST

tuberculin skin test

WAZ

weight-for-age z score

WHO

World Health Organization

Xpert

Xpert MTB/RIF