Abstract

Background: Chest radiography remains one of the most widely used imaging modalities for the detection of thoracic abnormalities, including pneumothorax and pulmonary masses. However, the frequency with which these findings are identified may vary according to demographic characteristics and radiographic acquisition techniques. Understanding such variability is important for interpreting large imaging datasets and may improve diagnostic assessment and epidemiological research. Therefore, this study aimed to evaluate sex-, age-, and projection-related differences in the dataset-derived detection frequencies of pneumothorax and pulmonary mass annotations in a large chest radiograph dataset. Methods: A retrospective cross-sectional analysis of 112,120 frontal chest radiographs obtained from 30,805 adult patients in the NIH ChestX-ray14 dataset was performed. Only anteroposterior (AP) and posteroanterior (PA) projections were included. Patients younger than 18 years, radiographs lacking demographic metadata, and lateral projections were excluded. Pneumothorax and pulmonary mass annotations were defined according to the corresponding NIH ChestX-ray14 labels generated from radiology reports using natural language processing. Patients were stratified according to sex and age group. Statistical analyses included chi-square tests, adjusted odds ratios (adjusted ORs), image-level multivariable logistic regression analyses, and corresponding patient-level generalized estimating equation (GEE) analyses to account for repeated radiographs from the same patient. Sensitivity analyses, including simulated annotation misclassification scenarios, were conducted to evaluate the robustness of the observed associations. Results: A total of 4988 pneumothorax and 5588 pulmonary mass annotations were identified. Pneumothorax detection frequencies were significantly higher among females aged 41–80 years (adjusted OR range: 0.66–0.68; p < 0.0001), whereas pulmonary mass annotations were more frequently detected in males, particularly in individuals aged 21–40 years (adjusted OR 1.52; p < 0.0001). Female patients were older on average and underwent PA imaging more frequently than male patients. AP projection was associated with lower pneumothorax annotation detection frequencies, whereas its association with pulmonary mass annotation detection was attenuated after patient-level analysis. Patient-level GEE analyses confirmed the robustness of the primary image-level findings. Conclusions: Detection frequencies of pneumothorax and pulmonary mass annotations varied systematically according to sex, age, and projection type, demonstrating substantial demographic and technical variability within a large chest radiograph dataset. These findings highlight the importance of considering demographic characteristics and imaging acquisition factors when interpreting large chest radiograph datasets. Because the findings were based on NLP-derived dataset annotations rather than reference-standard diagnoses, they should be interpreted as dataset-level associations rather than estimates of disease prevalence. Improved recognition of demographic and projection-related variability may contribute to more accurate interpretation of chest radiographs and more reliable imaging-based epidemiological research.

IPC Classification

Keywords