Individualized recurrence risk prediction as a challenge to conventional eligibility criteria in phase III adjuvant immunotherapy trials for renal cell carcinoma.

Abstract

e16501 Background: Adjuvant immunotherapy is recommended for high-risk patients with renal cell carcinoma (RCC) following surgery. However, inconsistent efficacy across phase III trials suggests that current clinicopathologic eligibility criteria may inadequately reflect the true risk of recurrence. We explored whether machine learning-based individualized risk prediction better discriminates outcomes than trial-defined eligibility in real-world patients with clear-cell RCC. Methods: Using the French national kidney cancer registry (UroCCR), we identified 2’403 patients with clear-cell RCC who underwent curative-intent surgery for non-metastatic disease or resected metastatic disease (M1 NED) prior to the approval of adjuvant pembrolizumab. Eligibility criteria from three phase III trials (KEYNOTE-564, IMmotion010, and CheckMate 914) were replicated to generate real-world trial cohorts. Individual recurrence risk was estimated with UroPredict 2, a machine-learning model derived from UroCCR data. Patients were stratified by predicted risk regardless of trial eligibility. Actual disease-free survival (DFS) and overall survival (OS) were estimated using Kaplan–Meier analyses and compared across risk strata with trial eligiblity criteria. Results: Marked variability in predicted recurrence risk was observed within both eligible and ineligible populations across all trial emulations. A substantial subset of patients meeting trial criteria were categorized as low risk (approximately 10% of eligible patients according to each trial eligibility criteria) and showed favorable DFS after surgery alone (actual DFS probabilities at 5 years above 81% for all criteria), consistent with limited expected benefit from adjuvant therapy. In contrast, more than 30% of patients excluded by each trial criteria were reclassified as intermediate or high risk by the predictor, and experienced poor oncologic outcomes (actual 5-year DFS probabilities of 57%–62%, depending on the trial eligibility criteria). Predicted risk groups consistently demonstrated clear separation of DFS and OS curves in each emulated trial cohort. Conclusions: Conventional eligibility criteria in phase III adjuvant immunotherapy trials incompletely identify patients at highest risk of recurrence after nephrectomy for RCC. Machine-learning-based individualized risk prediction may offer a more refined framework for selecting patients for adjuvant immunotherapy and for aligning treatment intensity with personalized recurrence risk.