Abstract Rationale The Radiographic Assessment of Lung Edema (RALE) score quantifies radiographic opacity extent and density in acute respiratory failure from chest X-rays (CXRs), correlating with clinical outcomes including mortality and ventilator-free days. While deep learning models enable automated RALE quantification with high reproducibility, their utility for longitudinal disease monitoring remains unexplored. We hypothesized automated RALE scoring could identify clinically meaningful temporal patterns in serial CXRs, providing scalable tools for tracking disease progression and treatment response in patients with pulmonary edema. Methods We trained three convolutional neural network (CNN) architectures (Siamese CNN, Quadrant Regression DenseNet, Quadrant Classification DenseNet) on 5,690 physician-annotated CXRs with quadrant-specific RALE scores, using standardized preprocessing including lung segmentation and brightness normalization. The best-performing model (Siamese CNN: ICC2,k=0.96, MSE=20.76) was applied to 1,000 patients from the NIH ChestX-ray14 dataset with ≥3 serial radiographs (14,982 total CXRs). Longitudinal consistency was assessed using intra-patient ICC3k. K-means clustering categorized predicted trajectories into severity groups: Mild (RALE 2-10), Moderate (10-24), and Severe (24-48). Within-patient variability was analyzed by plotting mean RALE versus standard deviation to identify stable versus progressive disease patterns. Results The model demonstrated strong longitudinal consistency (intra-patient ICC3k=0.70 overall; 0.82 for anteroposterior views), indicating reliable severity tracking within individual patients over time. Severity clustering revealed clinically interpretable separation with mean RALE scores of 4.1 (Mild), 16.6 (Moderate), and 31.4 (Severe). Variability analysis showed most patients maintained stable severity zones throughout follow-up, while a clinically relevant subset exhibited progressive radiographic worsening (Figure). These automated trajectories captured both persistent disease patterns and temporal evolution of pulmonary edema severity. Conclusion Deep learning-based automated RALE scoring enables scalable longitudinal monitoring of radiographic pulmonary edema across serial CXRs—a task requiring prohibitive manual effort for large patient cohorts. Automated severity trajectories demonstrated clinically meaningful stratification and temporal consistency, identifying both stable disease patterns and progressive deterioration. This approach provides objective, reproducible tools for monitoring treatment response, detecting early disease progression, and supporting clinical decision-making in critically ill patients. Future work should validate whether automated longitudinal RALE trends predict clinical outcomes including ICU length of stay, ventilator duration, and mortality. This abstract is funded by: None
Parvathaneni et al. (Fri,) studied this question.