Researchers have developed and validated PROGRESS (Prostate Cancer Global Risk Evaluation and Stratification Score), a simplified prognostic score for metastatic castration-resistant prostate cancer derived through machine learning–guided feature selection. Their findings, published in JCO Clinical Cancer Informatics, could inform clinical trial design and improve prognostic counseling.

The clinical heterogeneity and poor long-term survival associated with metastatic castration-resistant prostate cancer highlight the need for simpler prognostic tools that can reliably stratify patients by risk, according to senior study author Carlos M. Galmarini, MD, PhD, founder and CEO of Topazium Artificial Intelligence, Madrid, and colleagues.

“PROGRESS…enables effective mortality risk stratification of patients with [metastatic castration-resistant prostate cancer] using only three inexpensive, routinely available laboratory parameters,” the study authors wrote. “Developed through an integrative approach combining unsupervised [machine learning] with classical statistical modeling, PROGRESS consistently identified patient subgroups with markedly different survival outcomes across multiple independent cohorts.”

Study and Model Methods

The training cohort included 2,035 patients with metastatic castration-resistant prostate cancer treated with docetaxel plus prednisone across four clinical trials. A total of 159 baseline clinical and analytical variables were used to generate synthetic patient representations, which were analyzed using two popular machine learning techniques: Uniform Manifold Approximation and Projection (UMAP) for dimensionality reduction and Density-Based Spatial Clustering of Applications with Noise (DBSCAN) to identify outcome-related patient subpopulations.

The investigators used feature contribution analysis to select variables differentiating the subpopulations for further evaluation in a multivariable Cox proportional hazards regression model. This process yielded three independent predictors of outcome—prostate-specific antigen (PSA), alkaline phosphatase (ALP), and aspartate aminotransferase (AST)—that composed the PROGRESS score.

Overall survival was assessed using Kaplan-Meier analysis and Cox regression, and model discrimination was evaluated using the area under the receiver operating characteristic curve (ROC-AUC), time-dependent AUC, and Harrell’s concordance index (C-index). External validation involved two independent data sets comprising 1,239 patients across different treatment settings.

Key Findings

Two patient risk subpopulations were identified through unsupervised clustering using DBSCAN, with overall survival differing significantly between them (27.4 vs 17.7 months; hazard ratio [HR] = 2.20; P < .00001).

In the training cohort, PROGRESS demonstrated what the investigators described as strong discrimination (ROC-AUC = 0.89; time-dependent AUC = 0.60; Harrell's C-index = 0.69). Patients classified as increased risk by PROGRESS, with a prespecified cutoff (≥34), had significantly shorter median overall survival than those classified as low risk (18.3 vs 25.6 months; HR = 1.72; P < .0001).

PROGRESS appeared to demonstrate consistent survival prediction performance across all validation cohorts (all P < .0001). In the first validation set, the prognostic score yielded an ROC-AUC of 0.68, a time-dependent AUC of 0.65, and a Harrell's C-index of 0.69 for survival prediction, and achieved values of 0.64, 0.59, and 0.73, respectively, in the second validation set.  

“[PROGRESS’] ease of implementation and strong reproducibility support its potential utility as a pragmatic tool to complement existing prognostic frameworks in clinical practice,” the investigators concluded. “Future prospective studies and real-world validations will be essential to further define its role in guiding clinical decision-making and optimizing patient management in [metastatic castration-resistant prostate cancer].”

The study authors noted that generalizability of the prognostic score could be limited for broader real-world populations due to the retrospective nature of the data sets, and acknowledged that the model's performance degraded at longer time horizons, particularly at 30 months, suggesting the need for careful interpretation for longer-term outcomes. Additionally, the survival patterns used for training may not be reflective of newer treatment paradigms or improved supportive care measures. Despite these limitations, the study authors believe that PROGRESS is a clinically accessible pure prognostic tool for metastatic castration-resistant prostate cancer.

DISCLOSURES: Dr. Galmarini reported employment, a leadership role, stock and other ownership interests, and travel-related expense reimbursement from Topazium Artificial Intelligence; honoraria from BioPress; and consulting or advisory roles with Pharmactive. For full disclosures of the other study authors, visit ascopubs.org.

ASCO AI in Oncology is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (ASCO^®). The ideas and opinions expressed in ASCO AI in Oncology do not necessarily reflect those of Conexiant or ASCO. For more information, see Policies.