Clinical decision system for renal cell carcinoma integrating interpretable machine learning algorithms

Abstract

Background Kidney cancer is a highly heterogeneous oncologic disease with historically poor prognosis. Precise assessment of the risk of distal metastasis can facilitate risk stratification and improve prognosis for kidney cancer patients. Methods Data from the Surveillance, Epidemiology, and End Results (SEER) database, we identified 40,527 kidney cancer patients diagnosed between 2010 and 2017 were obtained. LASSO, univariate and multivariate logistic regression analyses were employed to screen independent risk factors for distal metastasis. Six machine learning (ML) algorithms including logistic regression (LR), Naïve Bayes Classifier (NBC), Decision Tree (DT), Random Forest (RF), Gradient boosting machine (GBM) and Extreme gradient boosting (XGB), were further applied to build the predictive models. After testing with ten-fold cross-validation and receiver operating characteristic (ROC) analysis, the model with the highest area under curve (AUC) was selected as the best performing model to establish the risk predictive nomogram and web calculator. Results In distal metastasis risk prediction, the XGB model had the best performance in both training (AUC = 0.91) and testing (AUC = 0.851) datasets among the six ML algorithms. Variables including marital status, sequence number, primary site, grade, pathological type, T-stage, N-stage, the calculated risk of XGB, surgical and radiation treatment were incorporated to establish a nomogram to predict the 1-, 3-, and 5-years survival probability. The calibration plots, decision curve analysis (DCA), ROC curves and Kaplan-Meier (KM) curves all verified the predictive utility of the nomogram. Conclusions We established a favorable prediction for the occurrence of distal metastasis with the ML model. The nomogram based on XGB algorithm can contribute to identify high-risk patients and provide optimal clinical strategies.