Construction of an interpretable machine learning clinical prediction model for renal fibrosis based on renal-cortex ultrasound radiomics

An interpretable clinical prediction model of renal fibrosis was constructed using grayscale ultrasonograms of the renal cortex and various machine learning algorithms. The study retrospectively collected 516 patients with chronic kidney disease (CKD) who visited the Department of Nephrology of Wuxi People’s Hospital Affiliated to Nanjing Medical University from April 2022 to June 2025. All patients underwent ultrasound-guided renal biopsy; renal fibrosis was confirmed histopathologically in 315 patients, whereas 201 patients had no renal fibrosis. The presence versus absence of fibrosis served as the binary outcome. The 516 CKD patients were divided into 414 training and 102 validation cases according to an 8:2 ratio. 3D-Slicer, the PyRadiomics package, and least absolute shrinkage and selection operator (LASSO) regression were used for manual labeling of the renal cortex, radiomic feature extraction, and dimensionality reduction within the training set only. Eight machine learning models were constructed, including LR, RF, GNB, MLP, SVC, XGBoost, LightGBM, and AdaBoost. The diagnostic performance of each model was evaluated through accuracy, precision, F1, sensitivity, specificity, and AUC with 95% confidence intervals. Hosmer-Lemeshow goodness-of-fit, calibration curves and decision-curve analysis were applied to assess the optimal model. SHAP (Shapley additive explanations) was used to interpret the radiomic features. A total of 837 radiomic features were extracted from the training set. Following intraclass correlation coefficient (ICC) analysis and LASSO regression, six robust radiomic features were ultimately selected: wavelet-LHH_glszm_SmallAreaHighGrayLevelEmphasis (F0), wavelet-LHH_glszm_Zone% (F1), wavelet-HLH_gldm_LargeDependenceLowGrayLevelEmphasis (F2), wavelet-LLL_glrlm_GrayLevelNonUniformity (F3), wavelet-LLL_glrlm_RunVariance (F4), and wavelet-LLL_glszm_GrayLevelVariance (F5). F0 demonstrated the strongest positive association, whereas F4 exhibited the most pronounced negative association with the target outcome (these are LASSO regression coefficients, not odds ratios). ROC analysis showed that the LightGBM model performed well. The accuracies of the training and validation sets were 0.803 and 0.692, respectively; the precision values were 0.798 and 0.724; F1 was 0.832 and 0.724; sensitivity was 0.870 and 0.724; specificity was 0.718 and 0.652; and the AUC was 0.887 (95% CI, 0.857–0.916) and 0.749 (95% CI, 0.657–0.843), respectively. The validation AUC of 0.749 indicates moderate discrimination, and the decline from the training AUC suggests a degree of overfitting. The prediction model of renal fibrosis based on LightGBM showed moderate-to-good discrimination and provided an interpretable imaging basis for the early screening of renal fibrosis. The selected feature set was not stable under bootstrap resampling and should be regarded as provisional. External, multi-centre validation is required before clinical application.