Exploratory associations between radiographic findings and metadata-derived proxies of 90-day follow-up in 112,120 ChestX-ray14 radiographs

Abstract

Abstract Chest radiography is widely used as an initial imaging modality. However, how specific findings relate to subsequent care or follow-up actions remains unclear. Prior studies have rarely examined follow-up actions, and potential sex-specific differences have been understudied. We analyzed 112,120 frontal chest radiographs from the NIH ChestX-ray14 dataset (63,340 male, 48,780 female). Images were labeled with 14 findings using a natural language processing (NLP) pipeline applied to reports. We modeled a metadata-derived proxy of 90-day follow-up using logistic regression, including sex-stratified analyses and interaction testing, with models adjusted for sex. Robustness was assessed through sensitivity analyses (30/60/180-day windows), patient-level clustering, and false discovery rate (FDR) adjustment. The strongest associations with proxy follow-up were observed for pulmonary edema (OR 10.6, 95% CI 8.5–13.2), pneumothorax (OR 7.6, 95% CI 6.7–8.6), and pleural effusion (OR 4.0, 95% CI 3.8–4.3). Interactions between sex and specific findings were modest but reached statistical significance for atelectasis ( P  = 0.003), pneumothorax ( P  = 0.0083), and emphysema ( P  = 0.0238). Radiographic findings were associated with metadata-derived proxy follow-up, but residual confounding from unmeasured factors (e.g., age, comorbidities, clinical context) likely remains. Sex-specific effects were statistically significant but small, and not clinically meaningful. Results should therefore be interpreted strictly as hypothesis-generating signals, not as causal evidence or clinically directive guidance.