Risk stratification using the Decipher 22-gene genomic classifier (GC) and digital pathology artificial intelligence (DPAI) in nearly 10,000 localized prostate cancer patients.

• DOI: 10.1200/jco.2025.43.5_suppl.408
• Publication Date: 2025-02-18T14:32:12Z
• Authors: Daniel Eidelberg ...; Soumyajit Roy; Ho Yin Ho; Xin Zhao; Alexander K. Haka...; Ahmad Alzein; Alboukadel Kassam...; Atallah Baydoun; Angela Y Jia; Phuoc T. Tran; Mamoudou Sano; Jason Hughes; Thomas Sbarrato; Jacques Fieschi; Elai Davicioni; Gerhardt Attard; Michael Leapman; Ashley Ross; Yang Liu; Edward M. Schaeffer
• Keywords: DECIPHER; Risk Stratification; Digital Pathology

408 Background: DPAI models have recently demonstrated the potential to improve risk stratification beyond routine clinical and pathologic variables. However, it is not known whether integrating information will enhance prognostic accuracy of validated gene expression tests. In this study, performance novel algorithms were assessed in context genomic classifier (GC). Methods: A prospectively collected cohort 9,874 patients with localized prostate cancer was retrieved from Veracyte GRID registry (NCT02609269). Scans at 40X magnification obtained for 17,701 H&E slides using an Aperio GT450 scanner (Leica). An open-source whole-slide pathology foundation model (GigaPath) used encode whole slide image (WSI) patches into digital features (DPIF). Attention-based multiple instance learning approach develop separate predict distant metastasis (DM) biopsy radical prostatectomy (RP) training subsets. WSI, baseline variables, GC scores linked tokenization (Datavant) real-world data (Clarivate). The primary endpoint study DM. Adjusted Hazard ratio (aHR) multivariable Cox regression (MVA) modeling 5-year area-under curve (AUC) estimates compare DPIF. Results: Median follow-up (n=6,705; 239 DM events) validation (n=3,169; 110 cohorts 6.5 years. 999 patients, predicted AUC 0.80 (95% CI, 0.70-0.90), which exceeded NCCN group alone (AUC 0.68, 0.56-0.80) 0.76, 0.66-0.86). MVA including age, NCCN, score, DPAI, only showed (aHR 1.23 [1.03, 1.47]) 1.22 [1.04, 1.43]) be significantly associated per 10%, both p<0.05). integrated clinicopathologic features, did discrimination 0.80, 0.70-0.90). RP 1,492 0.81 0.73-0.88). improved 0.84 (0.77- 0.91). MVA, 1.27 [1.13, 1.44 [1.23, 1.69]) significant predictors (both p<0.001). Conclusions: To our knowledge, largest assessing value adding above a clinical-genomic model. results suggest that combination these sources may further prognostication, use genomics negated Ongoing efforts larger are underway identify optimal scenarios utility decision making.