AI-Based Prostate Cancer Guideline Interpretation Shows Scaling Feasibility

Artificial intelligence (AI)–based guideline interpretation of NCCN guidelines for germline and somatic testing for prostate cancer in the real-world setting may be feasibly scaled to increase testing rates, according to findings from a retrospective chart review presented at the 2026 American Society of Clinical Oncology (ASCO) Annual Meeting.

Moreover, from 2023 to 2025, the period by which the study was conducted, rates of NCCN-indicated germline (P <.001) and somatic testing (P <.001) were inconsistently performed yet steadily increased for each successive year. Additionally, orders for testing outside of NCCN guidelines were placed for 4%, 9%, and 18% of patients in 2023, 2024, and 2025, respectively.

Regarding the accuracy of untrained AI for NCCN-indicated genetic testing, AI testing recommendation concordance was high for both germline and somatic testing. In 2023, a total of 259 charts were reviewed, and included 194 patients who were eligible for germline testing and 91 who were eligible for somatic testing. A total of 54 patients (28%) underwent germline testing, and 24 patients (26%) underwent somatic testing, with an AI concordance of 97% and 100%, respectively.

Moreover, in 2024, 255 charts were reviewed, with 195 and 86 patients eligible for germline or somatic testing, respectively. Germline testing and somatic testing occurred in 82 patients (42%) and 37 patients (43%), with a 100% AI concordance among each. For 2025, 177 charts were reviewed, identifying 170 and 70 patients eligible for germline and somatic evaluations. Testing was utilized among 107 patients (63%) and 56 patients (80%), respectively, with 100% AI concordance for both groups. Notably, AI risk stratification, which was assessable for 2024 and 2025, was 92% and 100% in each respective year.

Investigators noted that somatic testing was “recommended” among patients with metastatic disease per NCCN 2023 to 2025 germline testing numbers, and included patients with high-risk, very high-risk, and M1a-c disease.

“Planned electronic medical record [EMR] integration will enable real-time testing identification and scaling across diverse practice settings,” Mitchell Singstock, MD, internal medicine resident at the Huntsman Cancer Institute, wrote in the presentation with study coinvestigators regarding future directions for research. “Ongoing analyses seek to clarify which interventions contributed to increased testing rates over the [3-year] period.”

The retrospective chart review was conducted among patients with advanced prostate cancer treated at a large multi-site community oncology practice. Clinical data underwent manual review to assign disease stage, NCCN risk stratification, germline and somatic testing eligibility, and testing presence.

In 2023, ChatGPT 4.0, which at the time was an untrained model, was provided explicit rules-based NCCN guideline logic to determine testing eligibility. ChatGPT 5.2, used in 2024 and 2025, was provided the guideline text and asked to infer risk category and testing eligibility independent of clinician coaching.

In the review, structured inputs included age and prostate-specific antigen (PSA) levels at diagnosis, TNM stage, Gleason score, and grade group. The primary outcomes of the study were the observed annual germline and somatic testing rates, as well as AI concordance with manual review, which is currently considered the gold standard for review of patient charts.

The study authors conducted the review based on suboptimal real-world testing rates for prostate cancer, despite NCCN guidelines defining eligibility based on disease risk and stage. Moreover, they highlighted studies showing that approximately one-third of men with metastatic disease undergo genomic testing. To better improve testing rates, they evaluated prostate cancer care across 3 consecutive years to quantify testing utilization and assess AI performance for risk stratification and testing eligibility.

“The dataset that we built can be a testing ground for what we build next because these are hand-curated charts. If you want to use that as your gold standard, AI was proven to be pretty accurate,” David M. Waterhouse, MD, MPH, chair of the ASCO State of Oncology Care in America (SOCCA) Committee, director of Oncology Hematology Care (OHC)’s Early Phase Clinical Trials, and principal investigator of phase 1 trials at OHC, stated in an interview with CancerNetwork® regarding the review findings. “These were in domains that were structured data––we didn’t use natural language processing [NLP] or something similar, but that can be built, and we have the dataset to build it with.”

Reference

Singstock MD, Mendenhall M, Arnal S, Waterhouse DM. Real-world genetic testing patterns in prostate cancer assessed via artificial intelligence: implications for NCCN guideline implementation. J Clin Oncol. 2026;44(suppl 16):5022. doi:10.1200/JCO.2026.44.16_suppl.5022