# Biological Interpretable Machine Learning Model for Predicting Pathological Grading in Clear Cell Renal Cell Carcinoma Based on CT Urography Peritumoral Radiomics Features

## TL;DR

This study uses machine learning on CT scans to predict kidney cancer severity without invasive biopsies, offering a non-invasive alternative for preoperative risk assessment.

## Contribution

A novel machine learning model using peritumoral radiomics features from CT urography to predict ccRCC pathological grading with high accuracy.

## Key findings

- XGBoost model achieved high AUCs (0.95-0.92) across training and validation datasets for predicting ISUP grading.
- High-grade tumors showed enrichment in metabolic and immune-related biological pathways.
- The model outperformed other machine learning models and showed prognostic significance.

## Abstract

Background: The purpose of this study was to investigate the value of machine learning models for preoperative non-invasive prediction of International Society of Urological Pathology (ISUP) grading in clear cell renal cell carcinoma (ccRCC) based on CT urography (CTU)-related peritumoral area (PAT) radiomics features. Methods: We retrospectively analysed 328 ccRCC patients from our institution, along with an external validation cohort of 175 patients from The Cancer Genome Atlas. A total of 1218 radiomics features were extracted from contrast-enhanced CT images, with LASSO regression used to select the most predictive features. We employed four machine learning models, namely, Logistic Regression (LR), Multilayer Perceptron (MLP), Support Vector Machine (SVM), and Extreme Gradient Boosting (XGBoost), for training and evaluation using Receiver Operating Characteristic (ROC) analysis. The model performance was assessed in training, internal validation, and external validation sets. Results: The XGBoost model demonstrated consistently superior discriminative ability across all datasets, achieving AUCs of 0.95 (95% CI: 0.92–0.98) in the training set, 0.93 (95% CI: 0.89–0.96) in the internal validation set, and 0.92 (95% CI: 0.87–0.95) in the external validation set. The model significantly outperformed LR, MLP, and SVM (p < 0.001) and demonstrated prognostic value (Log-rank p = 0.018). Transcriptomic analysis of model-stratified groups revealed distinct biological signatures, with high-grade predictions showing significant enrichment in metabolic pathways (DPEP3/THRSP) and immune-related processes (lymphocyte-mediated immunity, MHC complex activity). These findings suggest that peritumoral imaging characteristics provide valuable biological insights into tumor aggressiveness. Conclusions: The machine learning models based on PAT radiomics features of CTU demonstrated significant value in the non-invasive preoperative prediction of ISUP grading for ccRCC, and the XGBoost modeling had the best predictive ability. This non-invasive approach may enhance preoperative risk stratification and guide clinical decision-making, reducing reliance on invasive biopsy procedures.