A Pilot Study Protocol for AI-Assisted Interpretation of Chest X-rays for Pulmonary Abnormalities in Uganda.

BACKGROUND

Timely access to chest X-ray (CXR) imaging and interpretation remains a practical challenge in routine pulmonary care in Uganda, particularly in settings with limited specialist availability and diagnostic capacity. This study aims to develop a structured, locally derived dataset of annotated CXR images linked with clinical metadata and to evaluate the feasibility of a machine learning model to support the diagnosis of pulmonary conditions. The algorithms developed using metadata acquired in this study will also help predict which patients should be referred for chest X-ray imaging.

METHODS

This pilot cross-sectional study will enroll 420 participants from Mbarara Regional Referral Hospital and Divine Mercy Hospital. Consecutive sampling will be used to recruit patients undergoing CXR for suspected pulmonary conditions, as well as individuals with normal findings. De-identified CXR images will be linked to standardized clinical metadata, including demographics, symptoms, examination findings, and imaging parameters. Images will be labeled by trained clinicians using standardized protocols. The dataset will be partitioned into training, validation, and test sets. Machine learning models, including convolutional neural networks and multimodal approaches integrating imaging and metadata, will be developed and evaluated using receiver operating characteristic-area under the curve (ROC-AUC), sensitivity, specificity, precision, recall, and F1-score.

RESULTS

The study is expected to produce a curated dataset of 420 annotated CXR images, including both normal and pathological findings. A pilot machine learning model for identifying pneumonia will be developed and internally validated. Additionally, a regression-based model is anticipated to explore patterns associated with CXR utilization in this clinical setting.

CONCLUSION

This study will establish a locally derived CXR dataset and assess the feasibility of machine-learning-based diagnostic support in pulmonary care. The findings will inform future model refinement, external validation, and potential integration into clinical workflows in similar settings.