Abstract
The task of predicting prostate-specific androgen (PSA) recurrence following radical prostatectomy is important for the surveillance of patients with prostate cancer. This regression problem is complicated by the fact that data is censored, and there is no standard measurement of error for censored data. This paper applies modified regression trees, called Censored Time Trees (cTT^TM), to predict time to PSA recurrence. In order to assess the performance of cTT^TM, we explored different error measurements, such as the concordance index, AUC (area under the Receiver Operating Characteristic curve), sensitivity and specificity, and average error. Bagging of Censored Time Trees^TM improves their performance. Bagging cTT^TM also performs slightly better than support vector regression and linear programming, both modified for censored data.