Image-only and multimodal AI digital pathology biomarkers to demonstrate risk stratification across standard prostate cancer management strategies.

Abstract

5023 Background: Reliable risk stratification across standard treatment pathways is essential for the clinical adoption of precision medicine biomarkers in localized prostate cancer. We previously developed and validated a multimodal artificial intelligence (MMAI) model that integrates digitized hematoxylin and eosin (H&E) prostate biopsy slides with clinical variables to prognosticate distant metastasis (DM) and prostate cancer specific mortality (PCSM). To evaluate the independent prognostic signal derived from routine histopathology alone without clinical variables, we built an Image-only artificial intelligence model. In this study, we validate the prognostic ability of the new Image-only and existing MMAI models in a multi-institutional cohort of nonmetastatic prostate cancer patients receiving guideline-concordant standard-of-care treatments. Methods: The Image-only model was trained to estimate the risk of 10-year distant metastasis (DM) without clinical inputs. Both the Image-only and MMAI models were evaluated in a cohort of men from 3 institutions with non-metastatic prostate cancer treated with active surveillance (AS), radical prostatectomy (RP), or radiation therapy (RT). The primary endpoint was 10-year risk of DM, with prostate cancer–specific mortality (PCSM) as a secondary endpoint. Prognostic associations were evaluated within each treatment subgroup using Fine–Gray models. Results: Among 886 patients with Image-only scores and 911 with MMAI scores, approximately 36% were managed with AS, 41% with RP, and 23% with RT. The Image-only model continuous raw scores were consistently associated with DM risk across treatment modalities, including AS (subdistribution hazard ratio [sHR] 2.38, p < 0.001), RP (1.99, p < 0.001), and RT (sHR 2.84, p < 0.001). Similarly, the Image-only model calibrated scores showed significant prognostic association with DM across all treatment groups. The MMAI showed consistent prognostic performance for DM using both calibrated and raw scores (raw scores: AS sHR 2.87, p < 0.001; RP sHR 2.12, p < 0.001; RT sHR 2.73, p < 0.001). Both Image-only and MMAI scores were significantly associated with PCSM despite low event rates. Conclusions: Both Image-only and MMAI biomarkers demonstrate consistent prognostic performance across standard prostate cancer management strategies, including AS, RP, and RT, supporting their utility for risk stratification regardless of ultimate treatment selection. These findings highlight the robustness of image-derived prognostic information and demonstrate that routinely available H&E pathology alone captures clinically meaningful risk information that generalizes across treatment contexts. Together, these results support the use of AI-based digital pathology biomarkers for prognostication in localized prostate cancer.