Prediction of Glomerular Filtration Rate Following Partial Nephrectomy for Localized Renal Cell Carcinoma with Different Machine Learning Techniques.

Background and

Objectives

Partial nephrectomy (PN) is the preferred option for treating localized cT1 renal cell carcinoma (RCC), as it preserves renal function in most patients and offers non-inferior oncological outcomes compared to radical nephrectomy. In this study, we aimed to construct a predictive model for estimating the glomerular filtration rate (GFR) at one year after PN in patients with RCC, using various machine learning techniques.

Methods

Retrospective data were collected from two academic centers, covering surgeries performed between 2010 and 2022. GFR was estimated using the Chronic Kidney Disease Epidemiology Collaboration 2021 (CKD-EPI) formula. Univariable linear regression (LR) was used to identify significant clinical predictors of 1-year postoperative GFR, followed by multivariable LR. The dataset was split into training and testing cohorts in a 70:30 ratio. Internal validation was performed on the test cohort, and various machine learning methods, including artificial neural networks (ANNs), support vector machines (SVMs), random forests (RFs), and XGBoost, were compared.

Results

Among 615 patients treated with PN, 415 had complete follow-up GFR data and were included in the analysis. Only 8.7% of patients experienced significant GFR loss (>30% decrease) at 1 year. Multivariable LR identified baseline GFR (Estimate: 0.76, p < 0.001), tumor diameter on imaging (Estimate: -1.65, p = 0.005), and Charlson Comorbidity Index (Estimate: -1.95, p < 0.001) as independent predictors of 1-year GFR (R2 = 0.67). A 10-fold cross-validation of the multivariable model yielded an R2 of 0.68. In the testing cohort, ANN, SVM, RF, and XGBoost did not outperform the LR model, with R2 values of 0.68, 0.66, 0.64, and 0.55, respectively.

Conclusions

Preoperative factors, including baseline GFR, tumor size on imaging, and Charlson Comorbidity Index, are effective predictors of GFR at 1 year following PN. Our study demonstrates that a conventional LR model based on preoperative variables provides acceptable accuracy for predicting GFR after PN and is not inferior to more complex machine learning techniques.