Independent blinded validation of an AI-based digital histology classifier for prostate cancer recurrence and metastasis risk prediction.

5098 Background: Artificial intelligence (AI) tools which identify pathology features from digitized whole slide images (WSI) of prostate cancer (CaP) generate data to predict risk of disease recurrence and metastasis. PathomIQ and ISMMS have developed an AI-enabled prognostic test, PATHOMIQPRAD, which predicts risk of biochemical recurrence (BCR) and distant metastasis (DM) using WSIs. The objective of this study was to evaluate the clinical validity of PATHOMIQPRAD using a retrospective clinical cohort at Cleveland Clinic. We also compared the test's performance to Decipher, an established genomic risk classifier. Methods: We conducted a retrospective PATHOMIQPRAD analysis of CaP WSIs of patients who underwent prostatectomy at Cleveland Clinic from 2009-2022 and did not receive any adjuvant therapy before BCR. 263 patients in the cohort received definitive treatment with radical prostatectomy and had a median follow-up of 50 months. Of these patients, 65 patients had BCR, and 14 patients developed DM as of last follow up. WSIs were de-identified, anonymized, and patient outcomes were blinded during the study. Patients were stratified into high-risk and low-risk categories based on pre-determined thresholds for PATHOMIQPRAD scores (0. 45 for BCR and 0. 55 for DM). The Kaplan-Meier method with log-rank was used to compare biochemical recurrence-free survival (BCRFS) and metastasis-free survival (MFS). Multivariable Cox proportional hazards regression was used to identify factors associated with BCR. Results: The rate of BCRFS and MFS were associated with both PATHOMIQPRAD score (BCR: >0. 45 vs. 0. 55 vs. 0. 6 vs. 0. 6 vs. <0. 6, p=0. 0095). All 14 patients who had DM during the follow up time had a high PATHOMIQPRAD score. Univariate analysis shows that PATHOMIQPRAD reliably identifies patients at risk of BCR (HR: 4. 19, p< 0. 0001), and had comparable prognostic performance to Decipher (HR: 2. 83, p=0. 0013). Multivariate analysis (Table) shows that there was an increased risk of BCR in both the high-risk PATHOMIQPRAD (HR: 3. 58, p=0. 0005) and Decipher (HR: 2. 20, p=0. 0159) groups relative to the low-risk groups, which suggests that combining the two tests may further improve risk stratification. Conclusions: These results show that PATHOMIQPRAD continues to demonstrate clinical validity in predicting risk of BCR and DM with favorable performance compared to a commonly used genomic classifier. PATHOMIQPRAD may identify patients for early treatment intensification, as well as inform clinical trial patient selection. Table: see text