Radiomic-Clinical XGBoost Model Distinguishes Pediatric TB from Pneumococcal Pneumonia on CT

Combined radiomic-clinical XGBoost model achieved AUC 0.897 on test set for separating pulmonary TB from pneumococcal pneumonia in children — outperforming clinical-only (AUC 0.784) and radiomic-only (AUC 0.865) models. Single-center, small n; not yet externally validated.

• Retrospective single-center study: 52 children with pulmonary TB and 80 with Streptococcus pneumoniae pneumonia; 7:3 train/test split; 1,023 CT radiomic features reduced to 6 via variance threshold, univariate analysis, and LASSO.
• Combined model training AUC was 0.981, suggesting possible overfitting; test AUC 0.897 still exceeded both component models (clinical 0.784, radiomic 0.865 on test set). SHAP waterfall plots provided feature-level interpretability.
• Key limitation: small sample (n=132), single center, retrospective design — external prospective validation in diverse populations needed before clinical adoption.

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