ROC optimization may improve risk stratification of prostate cancer patients

Rational treatment decision requires accurate projection of the clinical course of a patient. Current methods in clinical outcome analysis mostly focus on population data. We investigated the applicability and optimization of the widely used actuarial method to project individual clinical outcomes.

We designed and implemented a Clinical Outcome Prediction Expert (COPE) that performs, assesses, and optimizes actuarial prediction on individual cases. We analyzed a post-prostatectomy database, consisting of 1043 patients. Sixty percent of the database was used for training and 40% for validation. Stratified actuarial curves are used to project individual outcomes. The prostate-specific antigen (PSA) level, the Gleason score, and the clinical American Joint Commission on Cancer Staging T-stage before treatment were used as predictors. The area under the receiver operator characteristic (ROC) curve was used to measure predictive performance.

We obtained simple optimized stratification of pretreatment PSA level of 10 ng/mL or less, or more than 10 ng/mL; Gleason score of 6 or lower, or higher than 6; and clinical AJCC T-stage of T2a or lower, or higher. The optimized univariate risk scores were used to generate a multivariate score. After optimization, we found the higher risk group consisted of patients with PSA more than 10 ng/mL, or with PSA of 10 ng/mL or less and Gleason score higher than 6 and clinical AJCC T-stage higher than T2a. The optimized multivariate risk score has the highest ROC area of 0.77 among all predictors.

The best conditions to perform actuarial prediction on individual cases are not known a priori and require optimization. This study shows that ROC optimization simplifies risk stratification and may improve the accuracy of clinical outcome prediction.