An interpretable machine learning model based on habitat radiomics combined with deep learning for predicting the WHO/ISUP grade of patients with clear cell renal cell carcinoma

Abstract Background This study aims to explore spatial heterogeneity within tumors, establish and validate an interpretable machine learning model combining habitat radiomics and deep learning, and investigate its predictive value for the World Health Organization/International Society of Urological Pathology (WHO/ISUP) grading system for clear cell renal cell carcinoma (ccRCC). Methods A total of 646 patients participated in this retrospective study. Enhanced CT cortical phase images, clinical characteristics, and imaging features were collected. Habitat regions were generated using K-means clustering. Intra-tumor (Intra), habitat (Habitat), 2D and 2.5D deep learning (DL) models were developed. Independent predictive factors were identified through univariate and multivariate regression analysis, and a logistic regression (LR) classifier was integrated into a fusion model. Model performance was assessed using SHapley additive explainability (SHAP) analysis. Results Age, tumor size, and necrosis emerged as independent predictors. The habitat radiomics model demonstrated superior performance to the intratumoral and 2.5D models, with validation and test set AUCs of 0.854 (95% CI: 0.795–0.905) and 0.862 (95% CI: 0.785–0.915), respectively. The fusion model achieved optimal performance, yielding AUCs of 0.901 (95% CI: 0.850–0.948) and 0.913 (95% CI: 0.857–0.960) for the validation and test sets. Calibration curves confirmed high predictive accuracy, while decision curve analysis (DCA) revealed greater clinical utility for the fusion model. SHAP interpretability elucidated feature contributions to model predictions. Conclusions The fusion model significantly improves WHO/ISUP grade prediction in ccRCC. By enhancing interpretability through SHAP analysis, this approach offers a clinically valuable tool for preoperative assessment.