Birgitta Versluijs

67 HRCT after hematopoietic cell transplantation Establishment of the final diagnosis For each patient, an experienced pediatric hematologist who was unaware of the HRCT scores (10 years of experience in clinical hematopoietic cell transplantation) decided on the most probable final diagnosis. For this purpose, the physician had access to all avai- lable data including initial symptoms, physical examination findings, laboratory tests (including microbial tests from bronchoalveolar lavage), and imaging tests as well as response to therapy and follow-up examinations after the disease episode. In case of uncertainty, a second pediatric oncologist-immunologist (10 years of experience in cli- nical hematopoietic cell transplantation) was consulted and a consensus diagnosis was obtained. All patients were given a diagnosis: either nonallo-LS (consisting of infection and toxicity) or allo-LS (including idiopathic pneumonia syndrome and bronchiolitis obliterans syndrome). Both idiopathic pneumonia syndrome and bronchiolitis obliterans syndrome were clini- cally defined according to international criteria as acute respiratory symptoms, hypoxe- mia, abnormal radiographic findings, and restrictive physiology in the absence of infec- tion or heart failure. 18 We only further analyzed the HRCT score in scans obtained early after hematopoietic cell transplantation because in this phase it is both difficult and of utmost importance to discriminate between allo-LS and nonallo-LS. Statistical analysis Reproducibility of HRCT scoring between and within observers was assessed visually in scatter-plots with a line of identity and by using an intra-class correlation coefficient. An intraclass correlation coefficient between 0.6 and 0.8 represents moderate agreement and a value above 0.8 represents good agreement. To determine whether the severity of HRCT ab-normalities was different between allo-LS and nonallo-LS, a Mann-Whitney U test was done. The variables that showed statistical significance were then analyzed using logistic regression to identify their effect on the likelihood the patient had allo-LS. The regression coefficients were then transformed to integers according to their relative contributions to the risk, leading to a simplified weig- hed score. The different allo-scores and their value as a diagnostic test for allo-LS were analyzed using the ROC curve method. Accuracy of a test was measured by the AUC: an area greater than 0.8 was regarded as a good test. Finally we estimated sensitivity, specificity, positive predictive value, negative predictive value, and absolute percentage of correctly diagnosed patients for different cutoff values for the allo-score. We used Graphpad and SPSS 20.0 soft-ware for data analysis. Data are given as median unless indicated otherwise, and the significance level was set at p < 0.05. 4