A Robust Machine Learning Protocol for Prediction of Prostate Cancer Survival at Multiple Time-Horizons

Abstract

Prostate cancer is one of the leading causes of cancer death in men in Western societies. Predicting patients’ survival using clinical descriptors is important for stratification in the risk classes and selecting appropriate treatment. Current work is devoted to developing a robust Machine Learning (ML) protocol for predicting the survival of patients with metastatic castration-resistant prostate cancer. In particular, we aimed to identify relevant factors for survival at various time horizons. To this end, we built ML models for eight different predictive horizons, starting at three and up to forty-eight months. The model building involved the identification of informative variables with the help of the MultiDimensional Feature Selection (MDFS) algorithm, entire modelling procedure was performed in multiple repeats of cross-validation. We evaluated the application of 5 popular classification algorithms: Random Forest, XGBoost, logistic regression, k-NN and naive Bayes, for this task. Best modelling results for all time horizons were obtained with the help of Random Forest. Good prediction results and stable feature selection were obtained for six horizons, excluding the shortest and longest ones. The informative variables differ significantly for different predictive time horizons. Different factors affect survival rates over different periods, however, four clinical variables: ALP, LDH, HB and PSA, were relevant for all stable predictive horizons. The modelling procedure that involves computationally intensive multiple repeats of cross-validated modelling, allows for robust prediction of the relevant features and for much-improved estimation of uncertainty of results.