AI-Derived Biomarker Predicts Benefit of Long-Term ADT in Prostate Cancer

Andrew J. Armstrong, MD, ScM, professor of Medicine, Surgery, Pharmacology, and Cancer Biology at Duke University School of Medicine, and member at Duke Cancer Institute, discusses the methods behind a study that established a digital pathology biomarker for the benefit of androgen deprivation therapy (ADT) in localized high-risk prostate cancer.

According to Armstrong, he and his fellow investigators hypothesized that artificial intelligence (AI) tools could identify predictive biomarkers for long-term benefit from ADT and for higher risk of poor outcomes without intensified therapy. Using biopsy data from NRG cooperative group trials, they identified a biomarker based on digital pathology, as well as age, PSA, Gleason score, and T stage.

They validated the effectiveness of this AI-derived biomarker by testing it on the data from the NRG/RTOG 9202 (NCT00767286) trial, which established the benefit of long-term ADT in some men. The AI biomarker proved to be prognostic of distant metastases in the study’s patients and of prostate cancer-specific mortality for the duration of ADT, which were the RTOG 9202 study’s primary and secondary end points. This approach was able to find key features that were missed with clinical biomarkers, Armstrong says.

TRANSCRIPTION:

0:08 | We hypothesized that artificial intelligence tools leveraging digital pathology may identify predictive biomarkers that could identify men that had an excellent long-term outcomes, but with a short version of ADT. Likewise, there may be some intermediate-risk men that actually have high-risk pathology features that may need intensified therapy. To validate this, we used a large collection of phase 3 clinical trials through the NRG cooperative group. That's one of the great things about this study. It took digital pathology images from prostate biopsies from all over North America that were conducted as part of legacy trials going back 20 years.

We developed this biomarker using digital pathology, as well as clinical variables like age, PSA, Gleason score, T stage, and incorporated that into an AI tool, which became a biomarker that was then locked, and then validated externally in a prospective randomized phase 3 study called RTG 9202. We picked this trial because it was a positive study that changed practice many years ago. This taught us that long-term ADT did improve survival in some men, and overall, was a positive study for survival and changed our practice. What we found is that the AI biomarker which we validated learned from digital pathology key features that are missed with typical clinical biomarkers.