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

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