3379 - Benchmarking Multimodal AI-Predicted Metastasis and Prostate Cancer-Specific Mortality Risk Against STAR-CAP: A 20,000-Patient Commercial Experience
Presenter(s)
D. E. Spratt1, H. C. Huang2, R. T. Dess3, M. J. Zelefsky4, A. Briganti5, P. Carroll6, H. Nagar7, S. J. Freedland8, R. J. Karnes9, S. Roy10, A. Y. Jia11, E. M. Horwitz12, J. M. Michalski13, P. T. Tran14, B. J. Stish15, P. L. Nguyen16, H. M. Sandler17, G. Attard18, M. Schipper19, and Y. Sun20; 1University Hospitals Seidman Cancer Center, Case Western Reserve University, Cleveland, OH, 2ArteraAI, Los Altos, CA, 3Michigan Radiation Oncology Quality Consortium Coordinating Center, Ann Arbor, MI, 4NYU Langone Health, New York, NY, 5Vita Salute San Raffaele University, Milan, Italy, 6Department of Urology, University of California San Francisco, San Francisco, CA, 7Department of Radiation Oncology, Memorial Sloan Kettering Cancer Center, New York, NY, 8Department of Surgery, Section of Urology, Durham VA Health Care System, Durham, NC, 9Department of Urology, Mayo Clinic, Rochester, MN, 10Rush University Medical Centre, Chicago, IL, 11Department of Radiation Oncology, University Hospitals Cleveland Medical Center/ Seidman Cancer Center, Cleveland, OH, 12Department of Radiation Oncology, Fox Chase Cancer Center, Philadelphia, PA, 13Washington University School of Medicine, St. Louis, MO, 14Division of Radiation Oncology, The University of Texas MD Anderson Cancer Center, Houston, TX, 15Department of Radiation Oncology, Mayo Clinic, Rochester, MN, 16Mass General Brigham Cancer Institute, Boston, MA, 17Oregon Health and Science University, Portland, OR, 18The Institute of Cancer Research, London, United Kingdom, 19University of Michigan, Ann Arbor, MI, 20Case Western Reserve University School of Medicine, Cleveland, OH
Purpose/Objective(s): The commercial multimodal artificial intelligence (MMAI) digital pathology–based prognostic biomarker for biopsy specimens provides individualized 10-year distant metastasis (DM) risk estimates, categorized as MMAI Low (<3%), Intermediate (3–10%), and High (>10%). Herein, we report the performance of >20,000 commercial samples’ risk estimates with that derived from the multinational STAR-CAP clinical cohort.
Materials/Methods: Commercial real-world MMAI biopsy test results from May 2024 and January 2026 were utilized (n=20,069), and MMAI 10-year DM risk estimates were extracted. MMAI-predicted risks were benchmarked against observed outcomes in the STAR-CAP multicenter cohort (n=19,684). As the MMAI cohort was enriched for NCCN intermediate risk patients, and the STAR-CAP cohort differed in baseline clinical-pathologic composition, the MMAI cohort was supplemented with individual-level MMAI 10-yr DM risk estimates from published trials of men with non-metastatic prostate cancer (n=3,335) to broaden representation. Stratified resampling with replacement (1,000 iterations) was performed to match the baseline risk distributions of the two cohorts. After matching, MMAI-predicted 10-yr DM risks were compared with observed 10-yr DM risks in the STAR-CAP cohort. Because the STAR-CAP cohort underpins a validated PCSM risk prediction model, we also compared MMAI-predicted 10-yr PCSM rates with observed 10-yr PCSM outcomes in the STAR-CAP cohort.
Results: In the commercial MMAI cohort, NCCN low-risk patients were 82.3% MMAI low, with the remainder nearly all MMAI intermediate. NCCN intermediate-risk patients were 67.5% MMAI low, 30.4% MMAI intermediate, and 2.1% MMAI high. NCCN high-risk patients were 21.9% MMAI low, 51.1% MMAI intermediate, and 27.0% MMAI high. The 10-year predicted DM rates for the matched MMAI cohort were compared to those observed in STAR-CAP by NCCN risk group: NCCN low-risk was 2.4% vs 1.0%, intermediate-risk was 3.4% vs 4%, and high-risk was 12.9% vs. 13.5%. 10-year PCSM predictions by MMAI compared to observed in STAR-CAP by STAR-CAP stage groups were 1.0% vs 0.3% (stage IA), 1.3% vs 0.8% (IB), 1.6% vs 2.0% (IC), 4.1% vs 3.3% (IIA), 4.2% vs 4.4% (IIB), 6.2% vs 9.5% (IIC), and 8.6% vs 11.7% (IIIA). Limitations include underrepresentation of historical NCCN very low risk in the MMAI cohort, and fewer samples and more clinical heterogeneity in higher STAR-CAP stage groups precluding robust comparisons.
Conclusion: In the first report of >20,000 commercial MMAI tests, there was expected heterogeneity of MMAI risk within each NCCN risk group, yet the MMAI 10-year predicted DM and PCSM rates were generally similar to that observed in a large multinational cohort. Future work will focus on prospective validation of MMAI-guided clinical decision-making and optimal integration with treatment selection algorithms.