Multi-center validation of an artificial intelligence system for detection of COVID-19 on chest radiographs in symptomatic patients

Objectives While chest radiograph (CXR) is the first-line imaging investigation in patients with respiratory symptoms, differentiating COVID-19 from other respiratory infections on CXR remains challenging. We developed and validated an AI system for COVID-19 detection on presenting CXR. Methods A deep learning model (RadGenX), trained on 168,850 CXRs, was validated on a large international test set of presenting CXRs of symptomatic patients from 9 study sites (US, Italy, and Hong Kong SAR) and 2 public datasets from the US and Europe. Performance was measured by area under the receiver operator characteristic curve (AUC). Bootstrapped simulations were performed to assess performance across a range of potential COVID-19 disease prevalence values (3.33 to 33.3%). Comparison against international radiologists was performed on an independent test set of 852 cases. Results RadGenX achieved an AUC of 0.89 on 4-fold cross-validation and an AUC of 0.79 (95%CI 0.78–0.80) on an independent test cohort of 5,894 patients. Delong’s test showed statistical differences in model performance across patients from different regions ( p < 0.01), disease severity ( p < 0.001), gender ( p < 0.001), and age ( p = 0.03). Prevalence simulations showed the negative predictive value increases from 86.1% at 33.3% prevalence, to greater than 98.5% at any prevalence below 4.5%. Compared with radiologists, McNemar’s test showed the model has higher sensitivity ( p < 0.001) but lower specificity ( p < 0.001). Conclusion An AI model that predicts COVID-19 infection on CXR in symptomatic patients was validated on a large international cohort providing valuable context on testing and performance expectations for AI systems that perform COVID-19 prediction on CXR. Key Points • An AI model developed using CXRs to detect COVID-19 was validated in a large multi-center cohort of 5,894 patients from 9 prospectively recruited sites and 2 public datasets. • Differences in AI model performance were seen across region, disease severity, gender, and age. • Prevalence simulations on the international test set demonstrate the model’s NPV is greater than 98.5% at any prevalence below 4.5%.