Predicting Nephrectomy Risk in Patients with Renal Cancer Using Real-World Electronic Health Records.

Area of Science:

• Medical Informatics
• Oncology
• Machine Learning

Background:

• Nephrectomy is a key treatment for renal cancer.
• Accurate prediction of nephrectomy likelihood is crucial for clinical decision-making and preoperative planning.

Purpose of the Study:

• To evaluate machine learning (ML) models for predicting nephrectomy risk in patients with malignant renal tumors.
• To identify key predictors of nephrectomy using real-world electronic health record (EHR) data.

Main Methods:

• Utilized real-world EHR data from the UF Health Integrated Data Repository (IDR).
• Trained and validated ML models using demographic, clinical, and laboratory data prior to diagnosis.
• Employed Extreme Gradient Boosting (XGBoost) and SHapley Additive exPlanations (SHAP) for model evaluation and feature importance analysis.

Main Results:

• XGBoost model achieved a high performance with an F1 score of 0.638 and an Area Under the Curve (AUC) of 0.807.
• SHAP analysis identified HbA1C, serum creatinine, blood urea nitrogen (BUN), BUN-to-creatinine ratio, and glucose levels as significant predictors of nephrectomy risk.

Conclusions:

• ML models show significant potential for predicting nephrectomy risk in renal tumor patients.
• Real-world EHR data can be effectively leveraged for developing robust predictive models.
• Findings support the integration of ML in personalized care strategies for renal cancer management.