The main objective of the present study was to develop and evaluate adaptive synthetic sampling algorithm -based machine-learning model for estimating the likelihood of upstaging to pT3a in individuals with cT1 RCC.

We conducted a retrospective analysis of 1012 patients diagnosed with clinical T1 renal cell carcinoma and treated surgically at the Affiliated Hospital of Qingdao University from June 2016 to August 2021. After randomly assigning patients to a train set and a test set in a 7:3 ratio, using adaptive synthetic sampling algorithm addressed the issue of class imbalance. LR, LASSO and RFE were applied to select features. Then, DT, SVM, RF, XGBoost and MLP methods were used to predict upstaging. The performance of the methods was evaluated by accuracy, recall rate, and area under the curve value on the test. SHAP was used to aid in the interpretation and understanding of for the optimal model.

A total of 30 models were established, and through comparison, The LASSO-MLP model with the adaptive synthetic sampling algorithm train set achieved the best performance, which had the accuracy of 0.78, the AUC of 0.76 and the highest recall rate of 0.80. Compared to the original train set, the adaptive synthetic sampling algorithm train set improved the predictive performance in most of the models. SHAP analysis revealed that the tumor maximum diameter was the most important factor influencing upstaging, and other most selected features were relatively concentrated, demonstrating their value as important indicators influencing upstaging.

This study demonstrated the important role of machine learning could play in predicting staging to pT3a from cT1 RCC patients. The use of the adaptive synthetic sampling algorithm in the issue of upstaging could improve performance of models.