External Validation of an AI-Derived Model in Patients With High-Risk Localized Prostate Cancer Starting ADT

ArteraAI Prostate—a multimodal, artificial intelligence (AI)-derived prognostic test—was successfully validated in the phase 3 STAMPEDE trials and provided stronger prognostic associations than individual clinical variables, according to a presentation at the European Society for Medical Oncology Congress 2023.

Improved prognostication will allow for better targeting of treatment combinations for advanced prostate cancer. ArteraAI Prostate was recently developed for prognostication in localized prostate cancer.

Charles T. Parker, a clinical research fellow and doctoral student at the UCL Cancer Institute, and colleagues attempted to validate ArteraAI in advanced prostate cancer final data from 3 STAMPEDE trials. They used the model to digitize whole-scan images from new hematoxylin and eosin (H&E) core prostate biopsies, local Gleason score (GS), tumor stage (T), age, and pre-androgen deprivation therapy (ADT) serum prostate-specific antigen (PSA).

Fine-Gray and Cox regression adjusted for treatment allocation and cumulative incidence analyses were performed to evaluate associations with end points—as defined in STAMPEDE—for both continuous score (per standard deviation increase) and categorical (quartile, Q). Prostate cancer-specific mortality (PCSM) was the primary end point.

Among the 3879 patients recruited from October 2005 to January 2014 to ADT with or without radiotherapy alone, or with docetaxel and with or without zoledronic acid, or with abiraterone acetate and prednisone, 3725 consented to tissue analysis. Of those who consented, 2079 had H&E with sufficient tumor, and 1964 had all of GS, T, and pre-ADT PSA.

Dr. Parker and colleagues found that ArteraAI was strongly associated with PCSM (hazard ratio [HR], 1.67; 1.5-1.84; P<.001), overall survival (HR, 1.51; 1.4-1.63; P<.001), failure-free survival (HR, 1.48; 1.38-1.59; P<.001), and metastatic progression-free survival (HR, 1.59; 1.46-1.73; P<.001). Additionally, ArteraAI Q4 versus Q1-3 showed similar results for all end points, including PCSM (HR, 2.27; 1.95-2.65; P<.001).

Among component clinical variables, P<.001 for PCSM were PSA Q4 versus Q1-3 (HR, 1.8; 1.54-2.11), GS 8-10 versus ≤7 (HR, 1.64; 1.36-1.98), and T4 versus T1-2 (HR, 1.77; 1.36-2.32).

ArteraAI Q4 versus Q1-3 was found to have more PCSM events at 5 years: 16% (11-21) versus 6% (4-8) in high-risk localized disease and 58% (52-64) versus 40% (36-43) in metastatic disease. Notably, a similar trend was found for high-risk localized disease treated with ADT with or without radiotherapy and with abiraterone acetate and prednisone (18% [9-28] versus 2% [0-4]).

“ArteraAI was successfully validated in STAMPEDE with stronger prognostic associations than individual clinical variables,” researchers concluded. “The multimodal artificial intelligence model identified poor prognostic features in prostate biopsies from high-risk localized and metastatic prostate cancer.”