As survival rates for renal cell carcinoma (RCC) improve with advances in surgical technique and systemic therapy, the long-term consequences of nephrectomy are receiving increased clinical attention. Chronic kidney disease (CKD) is one of the most common and impactful complications, affecting quality of life and long-term cardiovascular health.
Nephrectomy, whether radical or partial, reduces functioning nephron mass and can precipitate or accelerate CKD in patients who already have compromised renal reserve. However, the risk is not uniform: some patients develop significant CKD while others maintain adequate renal function for years post-surgery.
Accurate preoperative prediction of CKD risk would allow surgeons and patients to make better-informed decisions about surgical approach (radical vs. partial nephrectomy), guide post-surgical monitoring intensity, and enable early nephroprotective interventions for high-risk individuals.
This study utilized the KORCC database, a multicenter Korean registry comprising 4,389 RCC patients from 8 hospitals. This large, real-world dataset provided diverse clinical profiles and ensured sufficient statistical power to train and validate machine learning models reliably.
Nine machine learning classifiers were developed and compared: logistic regression, decision tree, random forest, gradient boosting, AdaBoost, XGBoost, LightGBM, support vector machine (SVM), and k-nearest neighbors. Each model was trained on the same feature set derived from preoperative clinical variables available at the time of surgical planning.
The primary outcome was new-onset CKD following surgery, defined according to standard KDIGO criteria based on estimated glomerular filtration rate (eGFR) thresholds. Both training and validation sets were carefully constructed to reflect the realistic class imbalance present in clinical practice.
Among the nine algorithms tested, gradient boosting delivered the best overall performance with an AUROC of 0.826, sensitivity of 0.594, and specificity of 0.877. This combination of high specificity and moderate sensitivity reflects a clinically meaningful balance, correctly identifying the majority of patients who will not develop CKD while flagging a substantial proportion of those at risk.
Tree-based ensemble methods, including random forest, XGBoost, and LightGBM, all performed competitively, generally outperforming linear models such as logistic regression and simpler classifiers like decision trees and k-nearest neighbors. This suggests that CKD prediction involves complex non-linear interactions between clinical variables.
The AUROC of 0.826 compares favorably to existing clinical scoring systems for post-nephrectomy renal function prediction, indicating that machine learning approaches offer a meaningful improvement over traditional statistical models for this task.
SHapley Additive exPlanations (SHAP) values were used to decompose each model's predictions and identify the most influential features. SHAP provides a game-theoretically grounded attribution of each variable's contribution to individual predictions, offering both global feature importance rankings and patient-level explainability.
The top predictors identified by SHAP analysis were preoperative estimated glomerular filtration rate (eGFR), albumin level, tumor size, and serum calcium level. Preoperative eGFR was by far the strongest predictor, which is clinically intuitive: patients with already-compromised kidney function before surgery are at the highest risk of developing CKD afterward.
Albumin level as a predictor reflects nutritional and inflammatory status, both of which affect post-surgical renal recovery. Tumor size influences the amount of kidney parenchyma that must be removed, directly impacting residual renal function, while calcium levels may reflect paraneoplastic phenomena associated with more aggressive disease.
The machine learning model offers actionable preoperative risk stratification that can directly inform surgical planning. Patients identified as high-risk for CKD may benefit from nephron-sparing partial nephrectomy rather than radical nephrectomy, even when radical resection might seem technically easier.
High-risk patients identified preoperatively should also be referred for nephrology consultation before surgery, allowing optimized management of modifiable risk factors such as blood pressure, diabetes, and nutritional status. Early nephrology involvement can help establish a baseline and plan post-surgical monitoring protocols.
Patient counseling is another important application: patients with high predicted CKD risk can make better-informed decisions about surgical approach, lifestyle modifications, and long-term surveillance planning. Shared decision-making is enhanced when quantitative risk estimates are available rather than qualitative assessments.
This study demonstrates that machine learning models trained on routinely available preoperative clinical variables can predict post-nephrectomy CKD with clinically meaningful accuracy. The gradient boosting model's performance on a large multicenter dataset supports its potential for integration into clinical decision support systems.
The SHAP-based explainability framework is particularly valuable in the clinical setting, where black-box predictions are often mistrusted. By clearly showing which patient characteristics drive each risk estimate, SHAP bridges the gap between algorithmic prediction and clinical acceptance.
Future work should focus on prospective validation of the model, incorporation of additional biomarkers (genetic, urinary, or imaging-derived), and integration with electronic health record systems to enable automatic risk scoring at the point of surgical planning. Such tools could meaningfully improve long-term survivorship outcomes for the growing population of RCC survivors.