Upcoming event

Development of an individual postoperative prediction model for kidney cancer recurrence using machine learning (UroCCR study 120)

  • Margue G.,
  • Ferrer L.,
  • Etchepare G.,
  • Bensalah K.,
  • Mejean A.,
  • Roupret M.,
  • Doumerc N.,
  • Ingels A.,
  • Boissier R.,
  • Pignot G.,
  • Parier B.,
  • Paparel P.,
  • Waeckel T.,
  • Bigot P.,
  • Colin T.,
  • Bernhard J-C.

Introduction & Objectives

Kidney cancer is of increasing incidence worldwide. It is most often diagnosed at a localized stage where surgical management is the gold standard. Current available prognostic scores offer moderate predictive performance which leads to difficulties in establishing follow-up recommendations for patients after surgery and selecting patients who could benefit from adjuvant therapy. Our objective is to develop a model for individual prediction of the recurrence risk after surgery using machine learning (ML).

Materials & Methods

From the french research network on kidney cancer prospectively maintained database UroCCR (NCT 03293563), a cohort of patients undergoing surgery between May 2000 and January 2020 for a localized or locally advanced renal cell carcinoma (pT1-T4, N0, M0) was analyzed. Patients with a genetic cancer, a concomitant malignant disease or a chronic inflammatory disease were excluded. For each patient, clinical, biological, histological, and radiological data were collected. Participating sites were randomly assigned to the training or testing cohort with a 67/33 ratio of patients. Missing data were multiply imputed using the MICE algorithm and gradient boosted trees. Several ML algorithms were trained on the training data set and parameters of each algorithm were optimized using repeated cross-validation procedure (5x10folds). C-index at 5 years was used as optimization metric. The predictive performance of the algorithm was then evaluated on the test dataset using C-index and time-dependent AUC.


In total, 3255 patients were split in a training and test cohorts of 2172 and 1083 patients, respectively. The enclosed patients had tumors with a mean size of 4cm and 71% of ccRCC; 3.6% had locoregional recurrence, 7% had metastatic progression and 2.4% died. The median follow-up was 25 months. The best results in DFS prediction were obtained using a Cox PH model including 18 variables with a C-index of 0,75 and an AUC of 0,71 at 5 years. Comparatively, the C-index of the UISS and SSIGN scores were of 0.61 and 0.72. Shapley values graphs were generated to display the predicted DFS and the relative contribution of each feature leading to the personalized prediction.



ML applied to data from patients undergoing surgery for localized or locally advanced kidney cancer appears to provide good individual DFS predictions comparing to usual scores.