Explainable machine learning for predicting distant metastases in renal cell carcinoma patients: a population-based...

Area of Science:

• Oncology
• Machine Learning in Medicine
• Biostatistics

Background:

• Distant metastasis significantly impacts renal cell carcinoma (RCC) prognosis.
• Early prediction of metastasis is vital for tailored treatment strategies and enhanced patient outcomes.
• This research focused on creating and validating a clinical model for predicting distant metastasis in RCC.

Purpose of the Study:

• To establish and validate a robust clinical prediction model for distant metastasis in renal cell carcinoma (RCC) patients.
• To leverage machine learning for accurate metastasis risk assessment.
• To develop a tool for aiding clinical decision-making in RCC management.

Main Methods:

• Utilized machine learning algorithms, including Extreme Gradient Boosting (XGB), on a large dataset (51,566 patients) from the SEER database for model development.
• Externally validated the model using data from 726 RCC patients.
• Optimized hyperparameters via grid search and tenfold cross-validation; assessed performance using AUC, AUPRC, and calibration curves.
• Employed SHAP for model interpretability and developed a web-based calculator.

Main Results:

• The XGB model achieved high performance in the training set (AUC: 0.957, Accuracy: 0.898) and demonstrated strong external validation (AUC: 0.742, Accuracy: 0.904).
• Identified key independent risk factors including tumor size, grade, T-stage, N-stage, radiotherapy, chemotherapy, and surgery.
• A functional web-based calculator was created based on the validated XGB model.

Conclusions:

• An XGB model was successfully designed and validated for predicting distant metastasis risk in RCC patients.
• The developed model and associated calculator can assist clinicians in making informed decisions for RCC patient care.
• This approach offers a promising tool for early identification and management of metastasis in renal cell carcinoma.