Multimodal AI refines EAU risk stratification in patients undergoing radiotherapy for prostate cancer

Abstract The European Association of Urology (EAU) risk groups guide treatment decisions in prostate cancer using prostate-specific antigen (PSA), ISUP grade, and clinical stage, but do not incorporate histopathological image information. We evaluated a locked multimodal artificial intelligence (AI) model that integrates routinely available clinical variables (age, PSA, and ISUP grade) with deep-learning features extracted from diagnostic biopsy whole-slide images for predicting outcomes after radiotherapy (RT). The model was originally trained to predict biochemical recurrence (BCR) following radical prostatectomy and was applied without retraining to 353 patients from the population-based STHLM3 trial (ISRCTN84445406) treated with RT. Study endpoints included BCR (Phoenix definition), high-risk BCR (HR-BCR), and prostate cancer-specific mortality (PCSM). Model performance was assessed using Harrell’s C-index and cumulative incidence analyses and compared with EAU risk groups. The continuous AI score showed C-indices of 0.65 (95% CI 0.55–0.74) for BCR, 0.84 (0.72–0.95) for HR-BCR, and 0.81 (0.63–1.00) for PCSM, compared with 0.60, 0.82, and 0.80, respectively, for EAU risk groups. Although paired differences in C-index between AI and EAU were not statistically significant, adding the continuous AI score to EAU classification improved model fit for BCR, HR-BCR, and PCSM. Exploratory AI risk stratification classified 18.9% of EAU intermediate-risk patients as AI high-risk and 28.2% of EAU high-risk patients as AI low-risk. Within the EAU high-risk group, patients classified as AI low-risk ( n = 24) experienced no BCR events during 10 years of follow-up, compared with a 10-year cumulative incidence of 28.7% (95% CI 17.3–41.2%) among AI high-risk patients ( n = 61). These findings suggest that multimodal AI may refine EAU risk stratification by identifying distinct prognostic subgroups within established risk categories, although prospective validation is needed before its potential role in individualized treatment decision-making can be established.