Optimizing Chest X-ray Referral Using Clinical Metadata: Pulmonary Disease Patterns and Diagnostic Access in Southwestern Uganda

Abstract

Background: Chest X-ray (CXR) imaging is important for diagnosing pulmonary and cardiothoracic conditions, but timely access remains limited in many low- and middle-income countries. This study characterized radiographic abnormalities, examined associated clinical factors, evaluated exploratory clinical metadata-based prediction models, and assessed barriers to CXR utilization in southwestern Uganda. Methods: This facility-based observational pilot study included 422 adults undergoing chest radiography for suspected pulmonary or cardiothoracic disease at two healthcare facilities. Prospectively collected demographic, clinical, environmental, and healthcare-access data were linked to routine radiographer reports. Radiographer-reported pneumonia, pleural effusion, and cardiomegaly were summarized descriptively. Associated factors were examined using multivariable logistic regression with complete-case analysis. Exploratory model discrimination was assessed using receiver operating characteristic analysis, while post hoc Stage 1 simulations examined trade-offs between imaging-referral volume and case detection. Results: Complete radiographic-outcome classifications were available for 403 participants. Pneumonia was the most frequently reported abnormality (18.9%), followed by pleural effusion (9.7%) and cardiomegaly (5.5%). Increasing age was independently associated with pneumonia (adjusted odds ratio (aOR) 1.318 per 10-year increase; 95% confidence interval (CI) 1.149-1.523) and cardiomegaly (aOR 1.756 per 10-year increase; 95% CI 1.383-2.240), but not pleural effusion. Higher body mass index was associated with lower odds of pleural effusion and slightly higher odds of cardiomegaly. The exploratory models showed apparent in-sample discrimination for cardiomegaly (AUC 0.881; 95% CI 0.807-0.963), pneumonia (AUC 0.787; 95% CI 0.728-0.849), pleural effusion (AUC 0.699; 95% CI 0.601-0.788), and any reported abnormality (AUC 0.790; 95% CI 0.742-0.836). Effective access declined after accounting for personnel availability, affordability, and willingness to undergo imaging; only 17% of participants reported being both willing and able to complete the diagnostic pathway. The Stage 1 simulations illustrated trade-offs between referral volume and case detection under selected thresholds. Conclusions: Routinely obtainable clinical metadata may contain useful information for preliminary preimaging risk stratification. However, the reported AUCs represent apparent discrimination within the model-development sample, and the referral strategies were neither prospectively implemented nor clinically validated. The findings are exploratory and hypothesis-generating. Independent radiologist verification, validation of the report-classification procedure, model calibration, internal and external validation, and prospective workflow evaluation are required before the proposed approach can support patient-level referral decisions.