Selection and evaluation of preoperative systemic inflammatory response biomarkers model prior to cytoreductive nephrectomy using a machine-learning approach

Abstract

IntroductionThis study aimed to determine the prognostic value of a panel of SIR-biomarkers, relative to standard clinicopathological variables, to improve mRCC patient selection for cytoreductive nephrectomy (CN).Material and methodsA panel of preoperative SIR-biomarkers, including the albumin–globulin ratio (AGR), De Ritis ratio (DRR), and systemic immune-inflammation index (SII), was assessed in 613 patients treated with CN for mRCC. Patients were randomly divided into training and testing cohorts (65/35%). A machine learning-based variable selection approach (LASSO regression) was used for the fitting of the most informative, yet parsimonious multivariable models with respect to prognosis of cancer-specific survival (CSS). The discriminatory ability of the model was quantified using the C-index. After validation and calibration of the model, a nomogram was created, and decision curve analysis (DCA) was used to evaluate the clinical net benefit.ResultsSIR-biomarkers were selected by the machine-learning process to be of high discriminatory power during the fitting of the model. Low AGR remained significantly associated with CSS in both training (HR 1.40, 95% CI 1.07–1.82, p = 0.01) and testing (HR 1.78, 95% CI 1.26–2.51, p = 0.01) cohorts. High levels of SII (HR 1.51, 95% CI 1.10–2.08, p = 0.01) and DRR (HR 1.41, 95% CI 1.01–1.96, p = 0.04) were associated with CSS only in the testing cohort. The exclusion of the SIR-biomarkers for the prognosis of CSS did not result in a significant decrease in C-index (− 0.9%) for the training cohort, while the exclusion of SIR-biomarkers led to a reduction in C-index in the testing cohort (− 5.8%). However, SIR-biomarkers only marginally increased the discriminatory ability of the respective model in comparison to the standard model.ConclusionDespite the high discriminatory ability during the fitting of the model with machine-learning approach, the panel of readily available blood-based SIR-biomarkers failed to add a clinical benefit beyond the standard model.