Background/Objective: Chest radiography (CXR) is routinely used in the evaluation of respiratory disease; however, differentiating COVID-19 from other viral pneumonias on CXR remains challenging due to substantial radiographic overlap. In this study, a deep learning-based CXR classification model using a ResNet-50 architecture was developed to categorize images as normal, COVID-19, or non-COVID viral pneumonia, with emphasis on bias mitigation and external validation. Methods: Model training and internal validation were performed using harmonized publicly available datasets with patient-level stratified five-fold cross-validation, while generalizability was evaluated using an independent real-world institutional dataset from Adan Hospital, Kuwait, which was excluded from all training, validation, and hyperparameter tuning stages. Results: On the public validation dataset (n = 847), the model achieved an overall accuracy of 96.8% with balanced class-wise performance, whereas performance on the independent institutional dataset (n = 320) decreased to 93.7%, consistent with expected domain shift. Calibration analyses demonstrated well-aligned probabilistic estimates on validation data and acceptable calibration on institutional data. Negative predictive values remained high for normal and COVID-19 classes across datasets. Exploratory decision curve analysis demonstrated net benefit patterns for COVID-19 predictions under hypothetical threshold assumptions. Conclusions: These findings indicate that, when developed with explicit bias-mitigation strategies and evaluated using independent institutional data, deep learning-based CXR analysis may provide supportive, non-diagnostic decision signals for radiology triage workflows; however, prospective multicenter validation is required prior to clinical adoption.
Masoomi et al. (Thu,) studied this question.