A CTCs-based recurrence model for non-metastatic renal cell carcinoma: integrating machine learning and SHAP...

Postoperative recurrence remains a significant challenge in renal cell carcinoma (RCC). While circulating tumor cells (CTCs) have emerged as promising prognostic biomarkers, the predictive value of their subtypes (epithelial, hybrid, mesenchymal) and their dynamic changes over time for postoperative recurrence is not yet fully understood. This study aimed to analyze CTCs characteristics and develop a prognostic model for recurrence prediction.

We included 124 patients after RCC resection, collecting serial peripheral blood samples at regular intervals post-surgery to quantify CTCs subtypes using standardized enrichment and identification protocols. We extracted 54 variables, comprising 45 CTC characteristics and 9 clinical/pathological factors. Seven machine learning algorithms were trained to predict recurrence based on these features, with the SHAP (SHapley Additive exPlanations) framework applied to interpret the model.

Over a median follow-up of 41 months, 24 patients experienced recurrence, while 100 remained recurrence-free. The Random Forest model demonstrated superior performance, achieving a training AUC of 0.84 and a validation AUC of 0.77. SHAP analysis identified key predictors, including pT stage, tumor size, and changes in mesenchymal and hybrid CTC counts.