A study published in European Urology describes the development and validation of a multimodal artificial intelligence (AI)–derived digital pathology biomarker designed to predict distant metastasis in patients with biochemical recurrence (BCR) following radical prostatectomy.
BCR after prostatectomy represents a heterogeneous clinical setting, with variable outcomes following salvage radiotherapy and uncertainty regarding the benefit of adding hormone therapy. To improve risk stratification, the investigators developed a multimodal AI (MMAI) model that integrates digitized hematoxylin and eosin–stained prostatectomy slides with routine clinical variables, including pathologic grade group, pathologic T stage, pre–salvage radiotherapy prostate-specific antigen level, age, and surgical margin status.
The model was trained and validated using specimens and clinical data from two phase 3 NRG/RTOG trials (9601 and 0534). Validation was performed in a cohort of 533 patients with a median follow-up of approximately nine years. The MMAI score was independently associated with the development of distant metastasis after adjustment for clinical factors and treatment. At 10 years, the time-dependent area under the receiver operating characteristic curve for distant metastasis prediction was 0.74 for the MMAI model, compared with 0.68 for a clinical nomogram based on standard variables.
For interpretability, patients were also categorized into high- and low-risk groups using a predefined threshold. The 10-year cumulative incidence of distant metastasis was higher in the MMAI high-risk group (25 percent) than in the low-risk group (8.8 percent). Analyses stratified by treatment suggested differences in outcomes between patients receiving salvage radiotherapy alone and those receiving combined radiotherapy and hormone therapy, depending on MMAI risk classification.
From a diagnostic standpoint, the study demonstrates the use of archived prostatectomy specimens and centralized histopathology slides to derive prognostic information beyond conventional clinicopathologic features. Pathologist review of image regions emphasized by the model indicated that established adverse morphologic patterns, such as cribriform architecture, contributed to risk predictions.
The authors note that the findings are based on retrospective analyses of trial cohorts and that additional validation is ongoing. Prospective evaluation of the biomarker as a laboratory-developed test is currently underway to assess its potential role in post-prostatectomy risk assessment.
Newsletters
Receive the latest pathologist news, personalities, education, and career development – weekly to your inbox.
