Predicting Prostate Cancer Survival

In the final project for PUBH 8475 - Statistical Learning and Data Mining, my group of completed a section of the Prostate Cancer Dream Challenge, where the goal was to predict survival of metastatic castrate resistant prostate cancer patients in the comparator arms of clinical trials run by AstraZeneca, Celgene, Seno., and Memorial Sloan Kettering Cancer Center using various clinical covariates. Prediction endpoints of interest were survival at 12, 18, and 24 months. We used four methods:

• a naive random forest which did not account for censoring

• random survival forests

• elastic net-penalized Cox proportional hazards model

• gradient boosting with negative log of the Cox partial likelihood as the loss function

We found that random survival forests and the elastic net Cox model had the best prediction performance, and were comparable to some of the better submitted results in the actual challenge. Final paper and slides for the project can be found below.