Setting the Stage: AI Prognostication After Radical Prostatectomy

Despite decades of refinement in surgical technique and postoperative surveillance, predicting disease course after radical prostatectomy (RP) remains one of the most difficult problems in prostate cancer management. Approximately one-third of men who undergo RP will develop biochemical recurrence—defined by a detectable postoperative prostate-specific antigen level—and for this group, determining who requires further treatment, and when, represents a persistent source of clinical uncertainty. In this first segment of a 5-part series, R. Jeffrey Karnes, MD, the Dr. Anson Clark Professor of Urology at Mayo Clinic in Rochester, Minnesota, provides context for why improved prognostic tools are needed and what standards new AI-based models must meet before urologists can confidently incorporate them into practice.

Karnes notes that the field has seen increasing interest in AI and digital pathology tools across oncology, but cautions that the practical utility of any new model depends critically on the rigor of its validation. From both a clinical and research perspective, he emphasizes the importance of evaluating validation breadth—including the number of patients studied and the diversity of the populations represented—before drawing conclusions about generalizability. Variables such as race, geographic region, and socioeconomic factors can meaningfully influence prostate cancer biology and outcomes, and Karnes argues that external validation across varied populations is essential to ensuring a model's findings are not artifacts of the development cohort.

Karnes also addresses how AI-based prognostic models relate to established genomic classifiers, such as Decipher and Polaris, which have already been integrated into prostate cancer decision-making at many centers. He does not yet characterize the relationship between multimodal AI models and tissue-based genomic classifiers as definitively complementary or competing, noting that the data needed to make that determination—including head-to-head comparisons and studies examining whether the tools capture concordant or disparate prognostic signals—are still emerging. He expresses openness to the studies currently under way and frames the development of this evidence base as an important priority for the field.