Computational pathology–based classifier for predicting Gleason grade group upgrading on radical prostatectomy from diagnostic biopsies.

329 Background: Prostate biopsy is the gold standard for diagnosing and risk stratification of prostate cancer. As therapeutic options expand, detailed biopsy analysis becomes critical. However, discrepancies between biopsy samples and radical prostatectomy (RP) Gleason grade groups occur, with about 30% of patients experiencing an upgrade. This inconsistency can lead clinicians to underestimate disease severity, impacting patient management and outcomes. This study aimed to identify patients likely to experience a Gleason grade group upgrade (GGU) at RP by utilizing an AI-based computational pathology model that analyzes spatial architecture of lymphocytes from baseline diagnostic biopsies. Methods: We retrospectively analyzed whole-slide images from diagnostic biopsies of 128 patients who underwent RP (29 upgraded) at Helsinki University Hospital. Patients were randomly divided into training (D 1 ) and test (D 2 ) sets, comprising 89 and 39 patients, respectively. Nuclei segmentation and classification were performed using a deep learning model, followed by the extraction of 350 features representing the spatial interactions between lymphocytes and other nuclei. Feature selection and hyperparameter tuning were conducted on D 1 using cross-validated decision trees. The top five features and optimal parameters were then used to fit a final model GUP (Gleason Upgrade Predictor) on D 1 . GUP was subsequently used to predict GGU on D 2 . For comparative analysis, we developed models M c , M p and M cp using clinical variables (age, blood PSA level, PSA doubling time, T-stage, comorbidity index), data on proportion of Gleason grade patterns in the biopsies, and their combination. Finally, we integrated clinical variables into GUP (GUP c ) to evaluate whether this combination would enhance predictive performance and improve explainability. Results: The top five features measure lymphocyte cluster irregularity and their interactions with other cells. GUP achieved an AUC of 0.83 (±0.08), outperforming models trained exclusively on clinical data, Gleason pattern proportions, or their combination. Notably, GUP’s performance further improved when it was integrated with clinical variables (AUC 0.87 ± 0.1; Table). Conclusions: Our AI-driven computational pathology model shows strong predictive performance for GGU. Although external validation is still required, this model holds promise for improving the accuracy of Gleason group assessments at diagnosis. Model performance comparison. Model AUC ± std Balanced ACC GUP: Top 5 features of lymphocyte spatial architecture 0.83 ± 0.08 0.83 M c : Clinical variables 0.58* ± 0.08 0.58 M p : Pathological annotations of Gleason patterns 0.55* ± 0.10 0.52 M cp : M c + M p 0.72* ± 0.11 0.72 GUP c : GUP + clinical data 0.87 ± 0.10 0.87 *Indicates a statistically significant value at p < 0.05 comparing the model to GUP c .