Chest radiograph reading panel performance in a Bangladesh pneumococcal vaccine effectiveness study

INTRODUCTION: To evaluate WHO chest radiograph interpretation processes during a pneumococcal vaccine effectiveness study of children aged 3–35 months with suspected pneumonia in Sylhet, Bangladesh. METHODS: Eight physicians masked to all data were standardised to WHO methodology and interpreted chest radiographs between 2015 and 2017. Each radiograph was randomly assigned to two primary readers. If the primary readers were discordant for image interpretability or the presence or absence of primary endpoint pneumonia (PEP), then another randomly selected, masked reader adjudicated the image (arbitrator). If the arbitrator disagreed with both primary readers, or concluded no PEP, then a masked expert reader finalised the interpretation. The expert reader also conducted blinded quality control (QC) for 20% of randomly selected images. We evaluated agreement between primary readers and between the expert QC reading and the final panel interpretation using per cent agreement, unadjusted Cohen’s kappa, and a prevalence and bias-adjusted kappa. RESULTS: Among 9723 images, the panel classified 21.3% as PEP, 77.6% no PEP and 1.1% uninterpretable. Two primary readers agreed on interpretability for 98% of images (kappa, 0.25; prevalence and bias-adjusted kappa, 0.97). Among interpretable radiographs, primary readers agreed on the presence or absence of PEP in 79% of images (kappa, 0.35; adjusted kappa, 0.57). Expert QC readings agreed with final panel conclusions on the presence or absence of PEP for 92.9% of 1652 interpretable images (kappa, 0.75; adjusted kappa, 0.85). CONCLUSION: Primary reader performance and QC results suggest the panel effectively applied the WHO chest radiograph criteria for pneumonia.