Radiomics based on bone age radiographs combined with machine learning for predicting growth hormone deficiency in children

Background This study aimed to develop and validate a radiomics model based on bone age radiographs combined with machine learning for identifying pediatric GHD. Objective Growth hormone deficiency(GHD) is a common cause of childhood short stature, and non-invasive detection approaches for GHD remain lacking. To evaluate the diagnostic value of a machine learning–integrated radiomics model derived from left-hand bone age radiographs for identifying GHD in children, thereby providing a non-invasive and efficient screening approach for clinical practice. Methods A retrospective study included 271 children with short stature between June 2016 and November 2019. Participants were grouped as GHD (peak GH ≤ 10 ng/mL) or idiopathic short stature (ISS) group(>10 ng/mL), and split 7:3 into training and validation cohorts via stratified sampling. We manually segmented the distal radius as Region of Interest(ROI), extracting 1562 radiomic features. After Intraclass Correlation Coefficient filtering (ICC > 0.90) and dimensionality reduction with mRMR and LASSO, 18 features formed a radiomics signature. Eight machine learning models were built and evaluated by ROC curves, HL test and SHAP analysis. Results The XGBoost model exhibited the best diagnostic performance, with an AUC of 0.896 (95% CI: 0.886–0.979), accuracy of 81.5%, sensitivity of 71.4%, and specificity of 92.3% in the validation set. Its performance was significantly superior to that of the other models, including SVM (AUC = 0.841) and LGBM (AUC = 0.857). Among the selected radiomic features, first-order features accounted for the highest proportion, comprising 9 of the 18 retained features, which may be attributable to the two-dimensional imaging properties of bone age radiographs. Conclusion Radiomics based on left-hand bone age radiographs, when integrated with XGBoost machine learning, can effectively differentiate children with GHD from those with ISS, providing a new framework for a non-invasive and highly accurate auxiliary diagnostic tool in clinical practice.