Distant metastasis prediction in renal carcinoma: a retrospective cohort study using interpretable machine learning.

Background: Renal cell carcinoma (RCC) is the predominant subtype in renal carcinoma, characterized by its limited response to chemoradiotherapy. The development of distant metastasis greatly impacts patient prognosis. This study sought to identify risk factors associated with distant metastasis in RCC. Methods: Clinical data for RCC patients from 2010 to 2021 were extracted from the Surveillance, Epidemiology, and End Results (SEER) database. Univariate and multivariate logistic regressions (LRs) were employed to identify the most influential factors. Six machine learning techniques-LightGBM, LR, random forest, XGBoost, neural network, and gradient boosting machine (GBM)-were utilized to develop predictive models for distant metastasis. Shapley additive explanations (SHAP) analysis enabled visual interpretation for each patient. Model performance was evaluated through metrics such as accuracy, sensitivity, specificity, positive and negative predictive values, F1 score, area under the curve (AUC), Brier score, and estimated calibration index. Results: In predicting distant metastasis, the GBM model surpassed the performance of the other five machine learning algorithms. The GBM model demonstrated best overall performance with an AUC value of 0.907 (95% confidence interval: 0.899-0.914). The GBM model's calibration curve achieved a minimum Brier score of 0.047 (range, 0.042-0.052). SHAP analysis independently confirmed key clinical factors linked to metastasis risk in RCC patients. Conclusions: This study created a precise machine learning model to predict distant metastasis in RCC patients, offering clinicians valuable decision-support for personalized clinical management.