1640P - Development and validation of a machine learning-based risk model for metastatic disease in nmCRPC patients

Background

Nonmetastatic castration-resistant prostate cancer (nmCRPC) is a clinical challenge due to high progression rate to metastasis and mortality. To date, no prognostic model has been developed to predict the metastatic probability for nmCRPC patients.

Methods

A total of 2,716 nmCRPC patients were included in this study. The training and testing datasets were derived from the latest Phase III clinical trial SPARTAN and ARAMIS, respectively. Regarding metastasis-free survival (MFS) as the endpoint, we subjected 13 clinical features, including NHT application, baseline PSA level, PSADT, previous treatments received, Gleason score, race, and laboratory indicators, to 10 machine learning models and their combinations in order to predict metastasis. Model performance was assessed through accuracy (AUC), calibration (slope and intercept), and clinical utility (DCA). Risk score calculated by the model and risk factors base on 8 identified variates were used to metastatic risk stratification.

Results

The final prognostic model included eight prognostic factors, including NHT application, Gleason score, previous therapy (both surgery and radiotherapy, or neither), Race (White), PSADT, HGB, and lgPSA. The prognostic model resulted in a C-index of 0.764 (95% CI 0.740-0.787) in internal validation and relative good performance through tAUC (>0.70 at 3-month intervals between 6 and 39 months) in external validation. In risk score stratifying strategy, compared with low-risk group, the metastasis HRs for medium- and high-risk groups were 1.70 (95% CI 1.38-2.08) and 4.66 (95% CI 3.85-5.63); as for risk factor count, the HRs are 1.98 (95% CI 1.50-2.61) and 4.17 (95% CI 3.16-5.52), respectively.

Conclusions

In this study, we developed and validated a machine learning prognostic model to predict the risk of metastasis in nmCRPC patients. This model can assist in the risk stratification of nmCRPC patients, provide guidance for follow-up strategies, and aid in the selection of personalized treatment intensities.

Clinical trial identification

NCT01946204, NCT02200614.

Editorial acknowledgement

Legal entity responsible for the study

The authors.

Funding

Has not received any funding.

Disclosure

All authors have declared no conflicts of interest.