3199 - A Simple Two-Feature Model may Outperform Conventional Risk Classifiers for Metastasis in Prostate Cancer

Purpose/Objective(s): Prognostic models that rely on quantifiable features (such as Age or PSA value) may reduce misclassification relative to models with qualitative features (i.e., Gleason’s score), which are subject to human error. We previously demonstrated that the maximum tumor diameter (MTD) of the dominant prostate cancer lesion seen on multiparametric MRI, and PSA density (PSAd), are independently prognostic for metastasis. This study aims to determine if a prognostic model incorporating these two features perform as well as conventional risk classifiers used for treatment intensity decisions.

Materials/Methods: 498 Patients with known conventional risk classifications (NCCN, CAPRA) with multiparametric MRI reports were identified from a prospectively collected institutional database. MTD and prostate volume were retrospectively abstracted from these reports or measured by an expert from the source images when missing. Time to metastasis was anchored at the date of the biopsy preceding treatment. Prostate volume was determined using an ellipsoid calculation of the width, height, and length. Cox regression was used to model the effect of the covariates and conventional risk classifiers, and a log-hazard estimate for each individual was derived. The log-hazard estimates were fitted to a time-dependent receiver operating characteristic curve to determine AUC.

Results: The median age, PSAd and MTD was 67.3 years, 0.2 (ng/ml)/cc and 13mm, respectively. 192 (38.7%) patients received surgery, 286 (57.4%) received Radiation ± ADT, 8 (1.6%) received focal therapies, and 10 (2.0%) received ADT alone. The distribution by NCCN risk was: Low, 7%; Favorable Intermediate, 24.9%; Unfavorable Intermediate, 50.8%; High 14.5%, and Very High, 2.8%. The median follow-up time was 5 years. The bivariate model of PSAd and MTD had the highest 5-year AUC (0.771), which was substantially higher than the AUCs for NCCN risk (0.63), or CAPRA continuous (0.65) or categorical (0.64), Table 1.

Conclusion: These findings highlight that a simple, two-feature model (MTD and PSAd) can outperform widely used risk classifiers for predicting metastasis. By focusing on objective measurements, this model avoids potential human error from qualitative metrics and underscores the value of quantitative data in refining treatment decisions. Notably, the bivariate model achieved a higher AUC than NCCN or CAPRA classifications, reinforcing its potential clinical value. Validation in independent datasets will be required for clinical adoption.