Explainable machine learning to predict immunotherapy outcomes in metastatic renal cell carcinoma - Meet-URO 15-AI study

Sara Elena Rebuzzi, Vanja Miskovic, Giuseppe Fornarini, Sara Ferri, Sebastiano Buti, Matteo Piceni, Alessio Signori, Leonardo Provenzano, Giuseppe Luigi Banna, Pasquale Rescigno, Marco Maruzzo, Davide Bimbatti, Beatrice Ramella Pollone, Umberto Basso, Ugo De Giorgi, Paolo Pedrazzoli, Luca Galli, Paolo Andrea Zucali, Fabrizio Di Costanzo, Aruni Ghose, Alessandra Laura Giulia Pedrocchi, Arsela Prelaj

Abstract Metastatic renal cell carcinoma remains clinically challenging because of heterogeneous outcomes and limited predictive biomarkers for immunotherapy. We performed an explainable machine learning analysis using data from the multicenter retrospective Meet-URO 15 study, including 571 patients with metastatic renal cell carcinoma treated with second-line or later nivolumab. Clinical and inflammatory variables were used to develop classification models for disease control rate, progression-free survival at 3 and 9 months, and overall survival at 6, 18 and 24 months, as well as survival models for continuous progression-free and overall survival. Model performance was assessed using weighted F1-score for classification and concordance index for survival analysis, with interpretability provided through Shapley additive explanations. The best classification performance was observed for 6-month overall survival using a support vector machine model combined with minimum redundancy maximum relevance feature selection, achieving an F1-score of 0.81 on the test set and 0.77 in external validation. In survival analysis, random survival forest achieved a test-set concordance index of 0.68 for overall survival. Inflammatory indices, IMDC score, hemoglobin, lymphocytes and platelets consistently emerged as relevant prognostic features. These findings support explainable machine learning as a transparent approach to refine outcome prediction in immunotherapy-treated metastatic renal cell carcinoma.