Background: Radiologically-confirmed pneumonia (RCP) is a specific end-point used in trials of Pneumococcal Conjugate Vaccine (PCV) to estimate vaccine efficacy. However, chest radiograph (CXR) interpretation varies within and between readers. We measured the repeatability and reliability of paediatric CXR interpretation using percent agreement and Cohen's Kappa and the validity of field readings against expert review in a study of the impact of PCV on pneumonia.
Methods: CXRs were obtained from 2716 children admitted between 2006 and 2014 to Kilifi County Hospital, Kilifi, Kenya, with clinically-defined severe or very-severe pneumonia. Five clinicians and radiologists attended a three-day training course on CXR interpretation using a WHO standard. All CXRs were read once by two local primary readers. Discordant readings and 13% of concordant readings were arbitrated by a panel of three expert radiologists. To assess repeatability, a 5% median random sample was presented twice. Sensitivity and specificity of the primary readers' interpretations was estimated against the 'gold-standard' of the arbitrators' results.
Results: Of 2716 CXRs, 2 were uninterpretable and 159 were evaluated twice. The percent agreement and Kappa for RCP were 89% and 0.68 and ranged between 84-97% and 0.19-0.68, respectively, for all pathological findings. Intra-observer repeatability was similar to inter-observer reliability. Sensitivities of the primary readers to detect RCP were 69% and 73%; specificities were 96% and 95%.
Conclusion: Intra- and inter-observer agreements on interpretations of radiologically-confirmed pneumonia are fair to good. Reasonable sensitivity and high specificity make radiologically-confirmed pneumonia, determined in the field, a suitable measure of relative vaccine effectiveness.