Using Machine Learning to Predict Survival in Patients with Metastatic Castration-Resistant Prostate Cancer.

Abstract

Non-specific clinical biomarkers have been shown to help identify prognostic risks in cancer patients. However, the accuracy of prognostic biomarkers for predicting survival in patients with metastatic castration-resistant prostate cancer (mCRPC) still has space for improvement. This study aimed to predict 3-year survival in mCRPC patients by analyzing clinical and demographic features. A total of 664 patients with 41 clinical and demographic variables were evaluated. We utilized the class-weighted XGBoost algorithm to address class imbalance and improve the accuracy of outcome predictions. The model achieved an accuracy of 0.73, an AUC of 0.74, a recall, precision and F1 score value of 0.84, indicating a good ability to distinguish between patients who survived less than or more than 3 years. Our findings suggest that PSA, along with other non-specific biomarkers such as albumin and LDH, are significant predictors of survival in mCRPC patients and can be successfully used in machine learning algorithms to predict survival in mCRPC patients.