Disease-free survival in operated patients with nmRCC using machine learning.

4539 Background: In patients operated for a non-metastatic renal cell carcinoma (nmRCC), risk evaluation of disease recurrence enables the therapeutic strategy to be adapted after surgery. Usual risk group stratification does not fully exploit the power of machine learning (ML) applied to multimodal data. We developed a ML model capable of calculating the risk of recurrence after surgery for nmRCC patients. Methods: A multicentric cohort of patients undergoing surgery between May 2000 and January 2020 for a nmRCC (pT any, N any, M0) was analyzed from data collected prospectively in the French research network on kidney cancer UroCCR (ClinicalTrials.gov: NCT03293563). Using clinical, biological, histological and radiological features, we developed a signature of patient-specific DFS. Participating medical centers were randomly assigned to training or validation cohorts with a 2:1 ratio of patients. The ML signature managed missing data using multiple imputation based on gradient boosted trees and chained equations. Several ML models were trained in a repeated 5x10-fold cross-validation procedure to optimize the integrated Area Under the ROC Curve (iAUC) between 0.5- and 5-years following surgery. The performance of the selected ML signature was then evaluated on the validation cohort and compared with usual risk scores. Results: 3354 patients were randomly assigned to a training cohort (n=2239 from 10 centers) and a validation cohort (n=1115 from 13 centers). Our risk score achieved an iAUC of 0.81 (0.76, 0.85) on the validation cohort and significantly outperformed the GRANT ( p <0.001), SSIGN ( p =0.03), and UISS ( p <0.001) risk scores. The Leibovich-2018 score was only defined in 51% of the patients, due to incomplete data. Permutation-based feature importance and Shapley values were used to estimate the contribution of each feature. The stratification of patients into four risk groups achieved an iAUC of 0.78 (0.74, 0.83) on the validation cohort. 190 (17%) patients were stratified in the low-risk group, with an actual 5-year DFS higher than 98% and no death observed within this horizon window. 188 (17%) patients were stratified in the high-risk group, with an actual 5-year DFS of 48%. Conclusions: The machine learning model estimating individual disease recurrence risk outperformed the predictive performances of usual risk scores while handling incomplete data in predictors. This could help in identifying patients for whom surveillance could be reduced and those who could be candidates for adjuvant therapies. [Table: see text]