Abstract 6314: Development of g-rate based random forest machine learning model to predict overall survival for patients with metastatic prostate cancer

Abstract Metastatic prostate cancer (mPC) is among the most common causes of cancer-related death affecting men worldwide. Currently, predictive biomarkers and models are available to estimate a patient’s prognosis, but we lack reliable treatment assessment measures once a patient starts on a treatment. We have developed the g-rate method to measure treatment efficacy, which estimates tumor growth (g) and decay (d) rates using prostate-specific antigen (PSA) values while the patient is on treatment and have shown that g-rate has a robust inverse correlation with overall survival(OS) in patients with mPC. We have now developed a machine learning model based on g-rate and other clinical measures to predict the survival probability over time for individuals who are receiving treatment for mPC. We included variables of patient demographics, Charlson Comorbidity Index (CCI), mPC history (Gleason score, hormone sensitivity, line of treatment, PSA at diagnosis), variables available at the first visit of the start of treatment such as first PSA and first CBC at start of treatment, and g-rate derived from serial PSA values while on treatment. A random forest machine learning (ML) algorithm was employed to construct the survival prediction model. Harrel’s concordance index (C-index), Uno’s C-index, and time-dependent area under the receiver operating characteristic (ROC) curve were used for performance evaluation. Permutation importance was applied to assess variable contributions. Our experiment uses 18, 018 treatment records of 11, 849 unique US veteran patients to train the model. Besides the training data, 4, 501 treatment records of 2, 963 unique patients are used for model evaluation. The model achieved a remarkably high Harrel’s C-index of 0.76 and Uno’s C-index of 0.747, with 23.68% of treatment records censored. The time-dependent ROC curve averaged 0.83, demonstrating robust predictive capability. The permutation importance analysis identified g-rate as the most influential variable, followed by the first PSA and first Hgb at the start of treatment. Our g-rate based ML model highlights that it is possible to estimate treatment efficacy early via reliable survival probability prediction while on a treatment for mPC. Once validated, it can be used for both research and clinical purposes to make timely treatment escalation or de-escalation decisions. Citation Format: Mohamed Azzam, Harshraj Leuva, Mengxi Zhou, Benjamin A. Teply, Raymond C. Bergan, Susan Elaine Bates, Shibiao Wan, Antonio Tito Fojo, Jieqiong Wang. Development of g-rate based random forest machine learning model to predict overall survival for patients with metastatic prostate cancer [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 6314.