Multimodal AI fusion of chest radiographs and clinical data boosts prediction of severe pediatric pneumonia

A multimodal AI model fusing chest radiographs with lab and demographic data predicted severe community-acquired pneumonia in children with ROC-AUC 92.91% (vs 78.28% for chest radiograph alone), in a retrospective study of 3,964 patients.

• Single-center retrospective study of 3,964 pediatric CAP patients from Wuhan Children's Hospital.
• The optimal multimodal model achieved a PR-AUC of 46.68±4.50% and ROC-AUC of 92.91±0.64% (chest radiograph-only: PR-AUC 16.33±1.98%, ROC-AUC 78.28±1.48%).
• SHAP analysis showed model attention areas overlapped with radiologist-annotated lesions, supporting interpretability.

RadPigeon summaries are original and for information only. They are not clinical advice.