INTRODUCTION
Surgical intervention is the treatment of choice for localized renal cell carcinoma (RCC). In current guidelines, the European Association of Urology and the American Urological Association recommend partial nephrectomy (PN) as the preferred treatment for T1a tumors (≤4 cm) and as an alternative for T1b tumors (4-7 cm) whenever feasible [1-3].
Despite these guidelines, the decision between PN and radical nephrectomy (RN) remains complex, often requiring a balance between oncologic control and renal function preservation. While PN has preferred for its nephron-sparing benefits, RN is often chosen in patients with chronic kidney disease, high surgical risk, anticipated prolonged operation time, potential PN-associated morbidities, and most importantly when prioritizing oncological outcomes or survival benefits [4,5].
Meanwhile, machine learning (ML) introduces a novel approach to guiding clinical decisions regarding PN and RN by considering oncological and renal functional outcomes [6-8]. While current prediction models, even including traditional approaches, can individually predict oncological outcomes such as recurrence or renal function outcomes, none have simultaneously addressed both recurrence and renal function [1,4,6-10]. Moreover, few studies focus specifically on T1 renal tumors, which is the largest population in the contemporary cohort of RCC patients, presenting a significant clinical challenge.
In the current study, we aimed to present a prognostic model that leverages ML to simultaneously stratify 5-year recurrence probability and estimated glomerular filtration rate (eGFR) by surgical approach in patients with T1 renal tumors, using only preoperative variables.
MATERIALS AND METHODS
1. Patient Characteristics and Study Variables
After obtaining approval from the institutional ethics committee (IRB No. 2024-1217), data of 892 consecutive patients with T1N0M0 RCC who underwent RN or PN at Asan Medical Center between 2007 and 2019 was retrieved. Patients with single kidney, bilateral tumors, hereditary renal cancers involving multiple nephron-sparing procedures in the same kidney, histology other than primary adenocarcinoma of the kidney were excluded. Patients without preoperative 99mTc-diethylenetriaminepentaacetic acid renal scans or follow-up until 5 years after surgery were also excluded, leaving a total of 823 patients in the analytical cohort.
In the cohort, we reviewed patient demographics and clinical characteristics, including age, sex, body mass index, and comorbidities, along with tumor-specific parameters such as size, radiologic assessment of histologic subtype, and R.E.N.A.L. nephrometry score [11], as well as renal functionrelated data. Recurrence was defined as radiographic evidence of loco-regional recurrence or distant metastasis identified during the 5-year postoperative follow-up period. Renal functional outcome was assessed with eGFR calculated using CKD-EPI formula at postoperative 5 years [12].
2. Model Development
Model development incorporated preoperative variables regarded clinically significant based on established renal tumor surgery literature. Preprocessing was applied, and these variables were directly included in model training without univariate analysis or feature selection, given their established clinical relevance. To ensure the proportionality of RN and PN groups, the dataset was initially separated into RN and PN subgroups before stratification by recurrence. Each subgroup was then splitted into 80% training and 20% testing sets to maintain the recurrence rate of 2.4% be consistent across both sets to address class imbalance. The stratified RN and PN subsets were subsequently merged. K-fold cross-validation (k=5) was implemented during training to enhance evaluation robustness and model stability.
For performance assessment, 5 ML models—gradient boosting machine (GBM), extreme gradient boosting (XGBoost), general linear model (GLM), distributed random forest (DRF), and deep learning (DL)—were employed to predict eGFR and recurrence probabilities in clinical T1 renal tumor patients 5 years postoperatively, considering the 2 different surgical methods. GBM, XGBoost, and DRF are based on decision tree models. GBM combines many small decision trees that learn in stages, often achieving strong performance albeit risking overfitting if parameters are not carefully tuned. XGBoost is a faster, more advanced form of GBM that reduces overfitting with built-in regularization and speeds up training by using parallel processing to run different parts of the model at the same time. DRF trains many decision trees simultaneously on different data subsets, providing robust performance with simpler tuning. GLM links inputs to outputs through linear mathematical relationships, making it easy to interpret but often less capable of handling complex data. DL uses neural networks with multiple layers to automatically learn features, with this study specifically using a multilayer perceptron.
Model evaluation metrics included the area under the receiver operating characteristic curve (AUROC) to assess classification quality, the area under the precision-recall curve (AUPRC) to emphasize performance on the imbalanced dataset, and log-loss to measure overall predictive accuracy. Precision and recall were also analyzed to evaluate the models' ability to identify true positives and minimize false negatives, as accuracy alone can be misleading in imbalanced datasets. Data preprocessing, model training, and performance evaluation were conducted via Python version 3.11.8, with statistical significance determined at p<0.05.
RESULTS
The analytical cohort included 463 (56.3%) T1a cases, 75.6% of which underwent PN, and 360 (43.7%) T1b cases, 38.1% of which underwent PN (Table 1). Median age was 54 years (interquartile range, 46-62), with 68.9% being male. Histologic subtype was presumed to be of clear cell type in 82.7% of patients. Recurrence was observed in 1.1% of T1a and 4.2% of T1b patients at 5 years. The training set included 657 patients (79.8%) and the test set 166 (20.2%), maintaining similar distributions of surgery method and clinical stage (Table 2).
We developed a model including only the clinically significant preoperative parameters (Fig. 1). Among the models, the XGBoost demonstrated the highest performance in predicting 5-year recurrence, achieving an AUROC of 0.8850 and an AUPRC of 0.3790, indicating strong predictive accuracy (Table 3). Conversely, DL had the poorest performance, with the lowest AUROC (0.5679). GLM, GBM, and DRF produced intermediate outcomes, with AUROC values ranging from 0.6474 to 0.8488 (Fig. 2).
With respect to 5-year eGFR prediction, the GLM demonstrated the lowest RMSE in the test set (12.700), with respective R2 values of 0.688 and 0.694, indicating strong predictive accuracy (Table 4). The GBM and XGBoost models exhibited comparable RMSE and R² values; however, they showed larger differences between training and test set performance. Overall, GLM was the most accurate model for predicting 5-year eGFR (Fig. 2).
DISCUSSION
Our study demonstrates a novel application of ML to predict both oncological (recurrence) and renal function (eGFR) outcomes in patients with clinical T1 renal tumors undergoing nephrectomy. This dual-prediction framework bridges a significant gap in the literature, which has typically focused on either oncological or renal outcomes separately. Our findings emphasize the importance of preoperative factors such as tumor size, nephrometry score, and patient characteristics in prediction, which aligns with previous studies that identified these variables as critical in surgical outcomes [6,7,9,10]. Unlike prior models that often incorporated postoperative data, our approach relied exclusively on preoperative variables, thereby enhancing its relevance and applicability to real-world surgical decisionmaking.
For example, in our study, radiographic assessment of histologic subtype (clear cell vs. non-clear cell) was incorporated as a variable, which contributed to improving the accuracy of the recurrence prediction model. Despite potential limitations in the accuracy of radiographic histologic subtype assessment, we included this parameter given its clinical relevance; it is an essential factor that clinicians consider when determining the appropriate surgical approach [13].
The XGBoost model excelled in predicting 5-year recurrence, achieving the highest AUROC (0.8850) and AUPRC (0.3790) compared to other models. Precision-recall curves were employed to better assess model performance in this imbalanced setting, which provided more nuanced insights compared to traditional ROC curves [14,15]. High AUROC and AUPRC scores reinforce the reliability of the model in detecting true recurrence cases. The XGBoost model’s particularly high recall for recurrence prediction (0.0252) is notable given the low incidence rate (2.4%) in the cohort, with high recall being especially valuable to ensure true cases are not missed. Its strong recall shows its robustness in handling imbalanced data, especially when hyperparameters are fine-tuned to avoid overfitting [16].
On the other hand, the GLM exhibited the highest predictive accuracy for 5-year eGFR, as evidenced by the low RMSE in both training (10.831) and test (12.700) sets, and relatively strong R² values with 0.694 in the test set. This aligns with the nature of eGFR, which can often be estimated accurately through relatively linear associations with preoperative factors such as age, baseline renal function, and tumor size. Although some prior studies have employed binary classification for postoperative eGFR assessment [4,17], predicting eGFR as a continuous variable provides more precise information by offering exact estimates of renal function at 5 years, which is one of the goals of this study. The effectiveness of GLM in modeling straightforward linear correlations with continuous preoperative variables contributes to its predictive reliability in estimating renal function [18].
Several limitations need to be acknowledged in this study. First, its retrospective design may introduce biases related to patient selection and data availability. Additionally, the relatively small, single-institution sample limits cohort diversity, particularly regarding race, ethnicity, and regional healthcare practices. Another limitation is the lack of external validation; although our models performed well internally, their applicability to other clinical settings remains untested. Finally, the model does not account for intraoperative factors or postoperative complications that may affect long-term outcomes. Future research should include multi-institutional validation to better assess generalizability across varied populations and healthcare environments.
CONCLUSIONS
In this study, we developed and internally validated an online interface using ML to predict 5-year oncological and renal functional outcomes for patients with T1 renal tumors. Further multi-institutional validation is needed to confirm its generalizability and applicability across diverse clinical settings.
PDF Links
PubReader
ePub Link
XML Download
Full text via DOI
Download Citation
Print
