Comparison and interpretability of machine learning algorithms to predict survival of patients with prostate cancer.

OBJECTIVE: To evaluate the performance and interpretability of multiple ML algorithms for survival prediction in prostate cancer (CaP) and to assess whether these approaches can improve upon established clinical risk prediction models. METHODS: Using the Surveillance, Epidemiology, and End Results (SEER) 17 Database, we identified patients diagnosed with CaP from 2010 to 2017 and implemented a Bayesian Cox variable selection model to determine the most clinically relevant features. To predict overall survival (OS) and prostate cancer-specific survival (PCSS), we applied these features to train and test a traditional Cox proportional hazards (PH) model in comparison to ML models including XGBoost, random survival forests, and elastic nets as well as the Cancer of the Prostate Risk Assessment (CAPRA) score using 5-fold cross-validation. The models were repeated for cohorts limited to patients who underwent radical prostatectomy and those who had low-risk disease. RESULTS: For both OS and PCSS, XGBoost had the best area under the curve (AUC), concordance index, and Brier score, followed by elastic net, Cox PH, and RSFs. Feature analysis indicated that age was most predictive of OS, while cancer-specific characteristics such as Gleason grade group were the most important features for PCSS prediction. CONCLUSION: While the traditional Cox PH model remained a clinically robust cancer survival model, this study found that ML models demonstrated modestly improved CaP survival prediction and provided easily interpretable information on the factors most important for prediction. These findings underscore the ability of new ML algorithms to be clinically useful in predicting survival in patients with CaP, specifically by identifying high-risk patients who may require more intensive treatment plans and closer follow-up.