An outcome-blind H&E architecture score is associated with prostate cancer-specific mortality

Grade Group (GG) is the dominant prognostic variable in localized prostate cancer (PCa), but it does not encode the proportion of high-grade tumor or adverse subtypes such as cribriform growth. We developed an outcome-blind hematoxylin and eosin (H&E) architecture score from a Gleason-pattern classifier trained on SICAPv2 annotations (155 slides, 95 patients) with frozen Prov-GigaPath embeddings to generate patient-level architecture scores in two external radical prostatectomy cohorts. In PLCO (n = 1,086 patients, 66 prostate cancer deaths), the score was associated with prostate cancer-specific mortality (hazard ratio [HR] 2.57, 95% confidence interval [CI] 1.73-3.84) but not with death from other causes (HR 0.98, 95% CI 0.89-1.07), and it remained significant after clinical adjustment (HR 1.88, 95% CI 1.28-2.77). The association replicated for disease-free survival in TCGA-PRAD (n = 394 patients, 72 events; HR 2.83, 95% CI 1.64-4.88). The signal was consistent across different pathology foundation-model encoders and within the Gleason 7 stratum, and decomposed into high-grade and cribriform architecture. Transcriptomically, the score was specifically correlated with proliferation and genomic-instability programs, and its survival association in TCGA-PRAD was robust to tissue-source site. Among men with biopsy Gleason ≤6, the top score tertile had a higher 20-year prostate cancer mortality than the low-score group (4.20% versus 0.96%), flagging a candidate biopsy-occult higher-risk subgroup.