External validation of a pathology-based multimodal artificial intelligence biomarker for predicting prostate cancer outcomes after prostatectomy.

Abstract

5106 Background: Radical prostatectomy (RP) improves survival and delays metastasis in localized prostate cancer (PCa) patients (pts), yet 20-40% of men experience biochemical recurrence (BCR) within 10 years, with one-third of these progressing to metastatic disease. Predictive tools for risk stratification and treatment-decision making in this population remain limited. We previously developed and validated an RP digital pathology-based multimodal AI (MMAI) model using RP H&E images and select clinical variables to predict post-surgical outcomes in BCR pts (RP MMAI v1.1). We present the first external validation of this model in both BCR and non-BCR post-RP pts. Methods: Surgical pts with localized disease, clinical data, and long-term follow-up we re identified at UCSF (n=738). MMAI scores were generated from RP H&E images and clinical data (age, Gleason grade group (GG), pT-stage, surgical margins (SM), post-RP PSA). Fine & Gray regression with other-cause mortality as a competing risk was performed to determine the ability of MMAI score to predict any metastasis (primary analysis), bone metastasis (BM), and disease progression (DP: 2 consecutive PSA ≥0.2 ng/ml or salvage treatment) after RP. Hazard ratios, 95% confidence intervals, and p-values (for primary analysis) are reported. Results: MMAI scores were returned for 640 (87%) cases with images and clinical data. Median (IQR) post-RP follow-up was 11.5 (7.7-24.8) years. Post-RP Cancer of the Prostate Risk Assessment (CAPRA-S, range 0-12) scores were 56% low (0-2), 31% intermediate (3-5) and 13% high (≥6) risk. Characteristics at RP were 71% GG1/2, 64% pT2, and 79% negative SM. The majority of pts had undetectable PSA<0.05 after RP (87%). Cumulative incidence of DP and metastasis were 27% and 7% at 10 years, respectively. After adjusting for CAPRA-S, MMAI was independently associated with any metastasis (HR 1.76, [95% CI 1.23-2.53], p<0.001) and BM (HR 2.72, [95% CI 1.71-4.32]) in post-RP pts, as well as with DP in 561 pts with undetectable PSA after RP (HR 1.51, [95% CI 1.27-1.80]). Using a cutoff previously defined in BCR pts, 10-yr risk of any metastasis or BM after RP was higher in RP MMAI high risk (18% and 16%) vs. low risk pts (3% and 1%). In a subgroup of 211 salvage-eligible pts (detectable PSA and/or salvage treatment), MMAI remained independently associated with any metastasis (HR 1.71, [1.18-2.47]) and BM (HR 2.71, [1.74-4.23]) after CAPRA-S adjustment. Conclusions: This study validates the RP MMAI model, originally developed in BCR pts, as an independent prognostic tool in both BCR and general post-RP settings, even when controlling for a validated clinical risk model. These findings support its potential to guide personalized management strategies for post-RP pts, while offering advantages in accessibility, efficiency, and cost compared to existing platforms.