Supplement to

USING WHOLE BLOOD γ‐INTERFERON ASSAY TO IMPROVE DIAGNOSIS OF TUBERCULOUS PLEURAL EFFUSION

Subodh K. Katiyar1, Arun Sampath1,2, Shailesh Bihari1,2, Manju Mamtani2,3, Hemant Kulkarni2,3

1Department of Tuberculosis and Respiratory Diseases, Ganesh Shankar Vidyarthi Memorial Medical College, Kanpur, India 2Lata Medical Research Foundation, Nagpur, India 3University of Texas Health Science Center at San Antonio, Texas, USA

CONTENTS

Item Page Supplementary Note 1: Characteristics of the Study Subjects 2 Supplementary Note 2: Classification and regression tree analysis 3 Supplementary Figure 1: Diagnostic performance of continuous tests 4 Supplementary Table 1: Characteristics of the development sample 5 Supplementary Table 2: ROC analyses 6

1 Supplementary Note 1: Characteristics of the Study Subjects A. Development Sample The present study included patients with exudative pleural effusion who consecutively presented at the tertiary care center in the Department of Tuberculosis and Respiratory Diseases, Ganesh Shankar Vidyarthi Memorial Medical College (GSVMMC), Kanpur, India, between January 2006 and February 2007. Exclusion criteria were age <18 years, pregnancy, positive human immunodeficiency virus (HIV) serology and previous history of anti‐tuberculosis treatment. Thus all the study subjects were HIV‐seronegative, non‐tuberculosis contacts and not on immunosuppressive agents. Informed consent was obtained from all the patients. The study was approved by the Ethics Committee, GSVMMC, Kanpur, India. A total of 102 patients [ age 47.5 years, 72 (69%) males] were included. There were 38 (37.3%) confirmed TBPE patients based on pleural fluid culture positive and/or biopsy showing caseating granuloma; 14 (13.7%) probable TBPE patients; 44 (43.1%) non‐TBPE patients and 6 (5.8%) patients with pleural effusion of undetermined etiology (Figure 1). Pleural fluid culture was positive for M. tuberculosis in 18 TBPE patients (34.6%) of which eight patients were positive by pleural fluid culture alone. Out of the 18 culture positive patients only one had pleural fluid smear positive for AFB. Of the 50 patients with non‐TBPE; 26 patients (59.1%) had malignant pleural effusions; 15 patients (34.1%) had parapneumonic effusions; 2 patients (4.5%) had a rheumatoid etiology and one (2.3%) had uremia (Figure 1). The remaining six patients (5.8%) had pleural effusion of undetermined etiology. The malignant conditions observed were: Adenocarcinoma of lung (10), non‐Hodgkin’s lymphoma (3), breast cancer (3), Hodgkin’s lymphoma (2), small cell lung cancer (1), mesothelioma (1), ovarian cancer (1), gastric cancer (1) and unknown primary (4). The clinical characteristics of the study subjects are detailed in Supplementary Table 1. We observed that there was a statistically significant difference in the age and results of all the five tests studied for the TBPE versus non‐TBPE groups. B. Validation Sample We retrospectively collected a set of 84 cases of exudative pleural effusions reporting to the study center during the period August 1, 2008 to May 31, 2009. Of these 30 cases were excluded as per our study exclusion criteria: age <18 years (10 cases), pregnancy (2 cases), HIV sero‐positivity (2 cases) and previous history of anti‐ tuberculosis treatment (16 cases). The remaining 54 cases contained 25 cases of confirmed tuberculosis, 11 cases of parapneumonic effusions, 13 cases of malignant pleural effusions and 5 cases of undetermined etiology. Malignant conditions observed were: adenocarcinoma of lung (7), non‐Hodgkin’s lymphoma (2), breast cancer (2), ovarian cancer (1) and unknown primary (1). Excluding the five cases of undetermined etiology we used data from the remaining 49 exudative pleural effusions (25 TBPE and 24 non‐TBPE) for the purposes of validation. The validation sample contained 41 (83%) males and the mean (SE) age was 44.04 (2.31) years.

2 Supplementary Note 2. Classification and regression tree analysis

To identify a combination that would most parsimoniously and accurately classify the study subjects into TBPE and non‐TBPE we used the classification and regression tree (CART) analysis. CART is a deductive reasoning tool commonly used in data discovery to find relationships among a categorical outcome and a set of predictor variables [43‐46]. CART analysis comprises of four steps: i) Tree building: A classification tree is built using recursive splitting of nodes and each resulting node is ascribed to a predicted class; ii) Stopping rule: At this point a “maximal” tree is produced, likely overfitting the data; iii) Pruning: During this step, the maximal tree is trimmed into smaller informative candidate trees; and iv) Optimal tree selection: Tree that best fits (but does not overfit) the dataset is chosen. This approach is well suited for categorical outcomes. We chose the tree based on a series of binary diagnostic decisions that best fitted the data.

To build the tree, we pre‐set the maximum tree depth set to 4 splits and used 10‐fold cross‐validation as the method for pruning of potential trees. In the pruned optimum tree, we estimated the support (proportion of study subjects who could fulfill the classification rule for each terminal node) and confidence (proportion of subjects at a terminal node who were correctly classified). Using standard definitions, we then converted the information contained in the support and confidence into the likelihood ratio for diagnosis of TBPE for each terminal node. We did not force any predictor variable into the tree for either initial or subsequent splitting. Also, we assigned equal weight to all predictor variables. For statistical analyses we used the following three software programs: 10.0 (College Station, TX), CTree (http://www.geocities.com/adotsaha/CTree/CtreeinExcel.html) and OpenEpi (http://www.openepi.com).

3 Supplementary Figure 1. Diagnostic performance of the three tests (QFT‐G, ADA and L/N ratio) with continuous outcome. (A‐C) Box‐and‐whisker plots for the distribution of the indicated test results. In each plot, lighter shade indicates non‐TBPE patients and darker shade indicates TBPE patients. (D) Receiver operating characteristic (ROC) curve for the diagnostic performance of the color‐coded tests. Diagnostic accuracy is shown as the estimated area under the ROC curve (AUC) and its 95 % .

4 Supplementary Table 1: Characteristics of the study subjects in the development sample

Characteristic N* TBPE# Non‐TBPE P N = 52 N = 50 Age (mean ± SE yrs) 102 41.48±1.13 52.78±1.16 <0.0001 Males (n, %) 102 35 (67.31) 37 (74.00) 0.458‡ Rural area of residence (n, %) 102 44 (84.62) 38 (76.00) 0.273‡ Low socio‐economic status (n, %) 102 45 (86.54) 35 (70.00) 0.127‡ BCG received (n, %) 83 32 (68.09) 26 (72.22) 0.684‡ Tuberculin skin test positivity (n, %) 102 37 (71.15) 10 (20.00) <0.0001‡ QFT‐G (median, IQR IU/ml) 102 5.03, 11.18 0.22, 0.15 <0.0001 Pleural fluid analysis • PCR positivity (n, %) 102 32, 61.54% 3, 6.00% <0.0001‡ • ADA (median, IQR) 102 85.5, 51 26.5, 18 <0.0001 • Total cell count (median, IQR 89 1210, 1450 2345, 2530 0.023 cells/mm3) • PMNL cells (median, IQR %) 89 16, 9 20, 66 0.041 • Lymphocytes (median, IQR %) 89 79, 15 55, 47 <0.0001 • Lymphocyte/Neutrophil ratio (median, 89 4.47, 3.93 2.75, 5.41 0.0007 IQR) • Mononuclear cells (median, IQR 89 0, 4 7, 14 0.0002 cells/mm3) • Eosinophils (median, IQR cells/mm3) 89 0, 1 0, 1 0.364 • Mesothelial cells (median, IQR 89 0, 2 6, 10 <0.0001 cells/mm3) • Presence of atypical malignant cells (n, 89 0 (0.00) 10 (23.81) <0.0001 %) SE, standard error; IQR, interquartile range; IU, international units; U, units *Number of subjects on whom information was available #Includes confirmed and probable TBPE patients ‡Significance by Chi‐square test; all other significance values by Mann‐Whitney test

5 Supplementary Table 2. Comparative performance of tests to diagnose TBPE. Numbers are point and 95% confidence interval estimates for the indicated parameters.*

Test N Sensitivity (%) Specificity (%) LR+ LR‐ TST 102 71.2 (57.7 – 81.7) 80.0 (67.0 – 88.8) 3.56 (2.86 – 4.42) 0.36 (0.31 – 0.42) QFT‐G‐IT 102 90.4 (79.4 – 95.8) 86.0 (73.8 – 93.1) 6.46 (4.86 – 8.58) 0.11 (0.07 – 0.17) ADA 102 90.4 (79.4 – 95.8) 86.0 (73.8 – 93.1) 6.46 (4.86 – 8.58) 0.11 (0.07 – 0.17) L/N Ratio 89 85.1 (72.3 – 92.6) 50.0 (35.5 – 64.5) 1.70 (1.54 – 1.89) 0.30 (0.21 – 0.43) PCR 102 61.5 (48.0 – 73.5) 94.0 (83.8 – 97.9) 10.3 (5.14 – 20.5) 0.41 (0.37 – 0.45)

* Using the ROC analyses, the optimum cut‐off values for QFT‐G‐IT, ADA and L/N Ratio were 0.38 IU/ml, 38 U/L and 3.0, respectively.

6