Development and validation of a digital pathology artificial intelligence (DPAI)-derived risk score predicting Gleason grade group reclassification for patients who are candidates for active surveillance Authors: Brent Mabey, Lauren H. Lenz, Matthew J. Schiewer, Walter Rayford, Hassan Muhammad, Wei Huang, Robert Finch, Christina Nakamoto, Hosein Kouros-Mehr, Jeff Jasper, Hirak Basu, Chao Feng, Anurag Sharma, George Wilding, Rajat Roy, Dale Muzzey, Alexander Gutin Published: 2026 ABSTRACT ABSTRACTAimsActive surveillance (AS) allows selected men with localized prostate cancer to defer curative therapy and reduce treatment morbidity. Conversion from AS to treatment is commonly triggered by Gleason grade group (GGG) upgrading on confirmatory biopsy. We developed and validated a digital pathology artificial intelligence (DPAI)-derived risk score to predict GGG upgrading in AS-eligible patients.Materials and methodsThe DPAI model was trained using histopathology image features from diagnostic biopsies of 998 patients and validated in an independent cohort of 296 patients meeting criteria for AS. Logistic regression estimated the probability of confirmatory-biopsy GGG increase, and feature selection identified the most predictive variables.ResultsAI-GUR (Artificial Intelligence-Gleason Upgrade Risk) predicted GGG reclassification at confirmatory biopsy (OR 1.60; p = 0.0003) and provided information beyond conventional stratification (risk group, CAPRA) and cribriform morphology (all p < 0.01). Predicted risks were similar across time from diagnosis (~10–15% to ~85% at 1, 1.5, or 2 years; p for time = 0.50), consistent with initial biopsy mischaracterization rather than time-dependent progression.ConclusionsAI-GUR provides individualized estimates of confirmatory-biopsy GGG upgrading for AS candidates. Using DPAI may improve shared decision-making by complementing standard clinicopathologic tools and molecular testing using the same biopsy specimen, while informing the likelihood of grade upgrade at confirmation. Introduction Active surveillance (AS) allows patients with appropriately selected prostate cancer (PCa) risk profiles to safely avoid/delay morbidity associated with definitive therapy by closely monitoring disease and intervening upon progression. Despite strong long-term evidence supporting safety, AS remains underutilized in the United States, with only up to 60% of eligible patients electing this management strategy [1]. This gap persists largely due to uncertainty at the earliest decision points of care, where fear of undertreatment or missed opportunity for cure outweighs the objectively low risk of metastasis and PCa-specific mortality in patients with Gleason grade group 1 (GGG1) (low-risk) and select favorable GGG2 disease (intermediate-risk) [2]. Tools that improve individualized risk assessment and support shared decision-making across the key phases of AS–candidacy, confirmation of candidacy, and ongoing surveillance–have the potential to increase appropriate AS utilization while improving patient confidence in surveillance-based management. Identifying patients for AS candidacy largely relies on pathological assessment of diagnostic biopsies. National Comprehensive Cancer Network ® (NCCN®) Clinical Practice Guidelines in Oncology® [2] state that AS is the preferred management strategy for GGG1/low-risk disease and AS should be considered for GGG2/favorable intermediate-risk disease without unfavorable histology. Confirmation of AS candidacy includes pathological assessment of a follow-up biopsy, which allows for early detection of grade reclassification or cancer progression and mitigates any underestimates of tumor grade in the diagnostic biopsy due to baseline tumor heterogeneity, sampling variability, and subjective grading systems [2]. The following pathological considerations have been enumerated as rationale for transitioning from AS to definitive therapy: reclassification from GGG1 to ≥GGG2 in patients with low-risk disease; grade reclassification from GGG2 to higher, increase in percent Gleason Pattern 4 (GP4), increase in tumor volume, or rise in PSA density for patients with intermediate risk-disease [2]. Some variables captured coincident with diagnostic biopsy have been identified as prognostic for grade reclassification on confirmatory (i.e., first follow-up) biopsies, as described in the NCCN Guidelines®: tumor burden indicators (increased number positive cores, PSA density ≥0.15, faster PSA kinetics), tumor biology (increase in percentage GP4, presence of cribriform and/or intraductal carcinoma), and host/systemic factors (such as BMI) [2]. Importantly, existence of any one of these features alone is insufficient to require immediate use of definitive therapy but should be considered as part of a set of factors that inform the shared decision-making process in personalizing the intensity of an AS management protocol. While it is strongly recommended that all AS candidates undergo a confirmatory biopsy, the suggested timing ranges from 6 to 36 months, depending on a patient’s risk profile and multi-parametric MRI findings [2]. Further refinement of risk can be provided by advanced tools, such as the clinically validated CCR Score (Prolaris Prognostic Test; Myriad Genetics, Inc.) which identifies patients at higher risk of poor outcomes (metastasis and PCa mortality) from diagnostic biopsy material, even in patients with GGG1 and GGG2 disease [3]. The CCR Score has been commercially available in the US for over 10 years and is reimbursed by Medicare. The highest volume of test users are community urology practices. While the CCR Score refines baseline risk of adverse outcomes, it does not address the probability of grade reclassification on confirmatory biopsy, which is an orthogonal but clinically critical question. Current approaches do not allow for simultaneous refinement of initial risk as well as estimation of grade reclassification risk on confirmatory biopsy, limiting their utility in confirming AS candidacy. Here, we describe the development and validation of a novel DPAI-derived risk score designed to predict grade reclassification on confirmatory biopsy in patients with localized PCa who are candidates for AS. By integrating standard-of-care NCCN risk stratification and the CCR Score with a digital pathology artificial intelligence (DPAI) model trained on diagnostic biopsy material, this approach enables simultaneous assessment of AS candidacy and reclassification risk from a single biopsy specimen. This unified framework addresses a critical unmet need during AS confirmation, with the potential to inform the urgency of confirmatory biopsy, decrease uncertainty at early decision points, and support more confident, personalized shared decision-making. This strategy may help reduce unnecessary definitive treatment and promote more appropriate and durable implementation of AS in clinical practice. Materials and methods PATHOMIQ AI platform PATHOMIQ_PRAD is an AI-powered pipeline designed to generate AI-derived risk score that predicts biochemical recurrence and metastasis in localized PCa [4–6]. It comprises multiple modules including slide quality assessment, cancer detection, and cancer grading. The pipeline also produces detailed quantification of standard histopathologic features used in PCa risk assessment, including glandular-level Gleason grade percentages and ratios, tumor volume, and other metrics. Whole slide images of hematoxylin- and eosin-stained (H&E) slides were obtained via brightfield microscopy at 40x using the Aperio GT 450 digital slide scanner (Leica Biosystems). Images were managed with the Aperio eSLide Manager WebViewer (Leica Biosystems). The PRAD algorithm was applied to images within the HALO AP digital pathology platform (Indica Labs) in two steps as described in the Supplement. Histopathological tissue specimens Diagnostic biopsy samples were obtained during the routine execution of commercial laboratory testing. H&E slides were processed and scanned by a laboratory team who was blinded to outcome as well as to which slides would be used for training or validation. The training cohort consisted of specimens obtained from multiple independent cohorts as well as a clinically tested cohort, all of which came from both academic and community clinical practices. The foundation for the model presented herein arises from varied, generalizable sources. The validation cohort was similarly aggregated from multiple individual cohorts originating from both academic and community clinical practice. Finally, there was no overlap between training and validation cohorts. Specimens and associated data used in this study were de-identified in accordance with the HIPAA Privacy Rule (45 CFR §164.514) prior to analysis and did not include protected health information. Consequently, this activity does not constitute human subjects research as defined under 45 CFR §46.102(e) as there was also no interaction or intervention with individuals. As a result, the project was not subject to human subject research requirements, including IRB review and approval, and informed consent was not required. Training the AI-GUR score The AI Gleason Upgrade Risk (AI-GUR) score was trained on two observational cohorts, as well as a cohort of patients who had biopsies taken as part of their standard clinical care and had their diagnostic biopsy and follow-up biopsy sent for the CCR Score. The cohorts used in this study were not centrally reviewed. Training cohorts were restricted to patients considered candidates for AS by clinical features (diagnostic GGG1 or 2). The clinical cohort was enriched to have equal numbers of GGG1 and GGG2 patients, as well as equal numbers of events and nonevents. Events were defined as any increase in GGG (from GGG1 to ≥GGG2, and GGG2 to ≥GGG3). A summary of the full training cohort (n = 998) is given in Supplemental Table S1, with a patient flow diagram shown in Supplemental Figure S1. All training and validation analyses were performed in R version 4 or higher (R Foundation for Statistical Computing, 2026). AI-GUR was built by modeling the probability of GGG increase using logistic regression. Despite the timing between biopsies playing a role in GGG increase, logistic regression was used instead of survival analyses to account for the possibility of non-time-dependent factors in the events, such as resampling error and grading discordance by pathologist(s). Candidate variables for the model included 16-dimensional AI-embedding feature and AI-assessed histopathologic features produced by the PATHOMIQ_PRAD algorithm, the PRAD score (that combines the 16-dimensional AI-embedding and AI-assessed histopathologic features [4–6]), the molecular cell cycle progression score (CCP) [7] produced by the CCR Score, GGG at diagnosis (GGG1 vs GGG2), and interactions between diagnostic GGG and all other variables. AI-assessed histopathologic features measured as a proportion were logit transformed. Stepwise feature selection was used to choose candidate variables, starting from a model with all features included, then adding or dropping variables at each step based on their significance in a Chi-squared likelihood ratio test, with a cutoff at p = 0.05. After feature selection, the final model included two AI-assessed histopathologic features (proportion of identified cancer area with GP4, cribriform morphology, or perineural invasion, and the proportion of tissue area identified as tumor) as well as diagnostic Gleason and an interaction between diagnostic Gleason and proportion of tissue area identified as tumor. The 16-dimensional AI embeddings were included in candidate variable selection but were not included in the predictive model. Training and validation sets were identified prior to any results from the DPAI being released to the statistical analysis team and the validation set remained blinded until the model was locked after training. Prediction model development and the validation of prediction model function were reported in accordance with the Transparent Reporting of a multivariable model for Individual Prognosis Or Diagnosis (TRIPOD) guidelines updated for artificial intelligence (TRIPOD+AI) [8]. Statistical methods for validating the AI-GUR score Analyses followed a pre-defined statistical analysis plan, with non-pre-specified analyses labeled exploratory. Additionally, two statisticians performed the analyses independently to verify results. Inclusion criteria were pre-specified and included requiring diagnostic GGG1 or GGG2, being identified as a candidate for active surveillance by the test (CCR score ≤0.8 [3]), known time to confirmatory biopsy, and known GGG of confirmatory biopsy. Patients for whom a diagnostic biopsy slide was unavailable, or an AI-GUR score could not be calculated were excluded from the analysis set (Supplemental Figure S1). The endpoint of interest was any increase in GGG at confirmatory biopsy, referred to as GGG reclassification. Distributions of continuous variables were compared graphically, as well as with exploratory Kolmogorov–Smirnov tests and correlations. Rates were compared with tests of two proportions. The relationship between the AI-GUR score and other variables of interest and GGG upgrade status was evaluated using logistic regression models, accounting for cohort as a random effect. Results are reported as odds ratios (OR) per standard deviation (SD) for continuous variables, with 95% profile-likelihood confidence intervals and likelihood ratio test p-values. Significance was pre-defined at the alpha = 0.05 level. 3. Results Validation cohort characteristics The validation cohort included N = 296 patients from two prospective, observational studies and one retrospective, observational study (Supplemental Figure S1). The retrospective and one of the prospective studies were randomly split between training and validation (Supplemental Figure S1). A substantial majority of patients were NCCN low risk (81.8%) and GGG1 (84.8%). One third (99/296, 33.4%) of patients experienced GGG reclassification at confirmatory biopsy. The median time from diagnosis to confirmatory biopsy was 1.2 years (interquartile range 0.8 to 1.6 years; maximum 5.0 years; Supplemental Figure S2). A summary of the validation cohort is given in Table 1, including self-reported ancestry. Nearly 25% of the validation cohort self-reported having Black or African ancestry. Additionally, the validation cohort was well-represented for baseline risk profiles suitable for AS candidacy, consisting of 81.8% NCCN low-risk, 17.6% NCCN favorable intermediate-risk, and <1% NCCN unfavorable intermediate-risk patients (Table 1).Table 1.Characteristics of the validation cohort, N = 296. N (%) or Median (IQR)Age at diagnosis65 (60, 69)Self-reported ancestry White or European203 (68.6%) Black or African67 (22.6%) Other or multiple4 (1.4%) Unknown or not reported22 (7.4%)Diagnostic Gleason Score 3 + 3251 (84.8%) 3 + 445 (15.2%)Diagnostic PSA (ng/mL)5.1 (4.3, 6.3)Percent of biopsy cores positive for cancer21% (14%, 33%)Clinical T stage T1275 (92.9%) T221 (7.1%)NCCN Risk Category Low242 (81.8%) Favorable Intermediate52 (17.6%) Unfavorable Intermediate2 (0.7%) High0 (0%)Years to first follow-up biopsy1.2 (0.8, 1.6)Gleason pattern 4 cribriform identified by AI72 (24.3%)AI-GUR Score−1.264 (−1.859, −0.912)Upgrade at first follow-up biopsy99 (33.4%) Table 1.Characteristics of the validation cohort, N = 296. N (%) or Median (IQR)Age at diagnosis65 (60, 69)Self-reported ancestry White or European203 (68.6%) Black or African67 (22.6%) Other or multiple4 (1.4%) Unknown or not reported22 (7.4%)Diagnostic Gleason Score 3 + 3251 (84.8%) 3 + 445 (15.2%)Diagnostic PSA (ng/mL)5.1 (4.3, 6.3)Percent of biopsy cores positive for cancer21% (14%, 33%)Clinical T stage T1275 (92.9%) T221 (7.1%)NCCN Risk Category Low242 (81.8%) Favorable Intermediate52 (17.6%) Unfavorable Intermediate2 (0.7%) High0 (0%)Years to first follow-up biopsy1.2 (0.8, 1.6)Gleason pattern 4 cribriform identified by AI72 (24.3%)AI-GUR Score−1.264 (−1.859, −0.912)Upgrade at first follow-up biopsy99 (33.4%) Characteristics of the validation cohort, N = 296. Validation of DPAI Gleason upgrade predictive score (AI-GUR) In a model with AI-GUR score and time between diagnosis and confirmatory biopsy predicting GGG upgrade at confirmatory biopsy, AI-GUR score was a statistically significant predictor (OR 1.60, 95% CI 1.23–2.11, p = 0.0003), whereas time (in years) since diagnosis was not (OR 1.09, 95% CI 0.85–1.38, p = 0.50; Table 2). The probability of GGG reclassification at first follow-up biopsy vs. AI-GUR score is shown in Figure 1, which demonstrates that with increasing AI-GUR score the probability of GGG reclassification increases similarly (from ~10% to ~85%) irrespective of time elapsed between biopsies (1, 1.5, or 2 years). Figure 1.Probability of Gleason upgrade at first follow-up biopsy vs AI-GUR for follow-up biopsies at 1, 1.5, and 2 years post-diagnosis.Horizontal axis label: AI-GUR Score; range negative 4 to 3 with ticks at negative 4, negative 3, negative 2, negative 1, 0, 1, 2, 3. Vertical axis label: Probability of upgrade at first follow-up biopsy; range 0.0 to 1.0 with ticks at 0.0, 0.2, 0.4, 0.6, 0.8, 1.0. Legend text: Years from diagnostic to first follow-up biopsy with series 1, 1.5, 2. Series 1: (negative 4, 0.10), (negative 3, 0.16), (negative 2, 0.25), (negative 1, 0.38), (0, 0.52), (1, 0.66), (2, 0.79), (3, 0.87). Series 1.5: (negative 4, 0.10), (negative 3, 0.16), (negative 2, 0.26), (negative 1, 0.39), (0, 0.53), (1, 0.67), (2, 0.80), (3, 0.88). Series 2: (negative 4, 0.11), (negative 3, 0.17), (negative 2, 0.26), (negative 1, 0.39), (0, 0.54), (1, 0.68), (2, 0.80), (3, 0.88).A graph of the sigmoid function showing probability of upgrade at first follow-up biopsy versus AI-GUR score.Table 2.Results of a multivariate logistic regression model predicting Gleason upgrade at first follow-up biopsy. Model includes a random effect for study.PredictorOdds Ratio per standard deviation (95% confidence interval)p-valueAI-GUR1.60 (1.23, 2.11)0.0003Years since diagnosis1.09 (0.85, 1.38)0.50 Probability of Gleason upgrade at first follow-up biopsy vs AI-GUR for follow-up biopsies at 1, 1.5, and 2 years post-diagnosis. Table 2.Results of a multivariate logistic regression model predicting Gleason upgrade at first follow-up biopsy. Model includes a random effect for study.PredictorOdds Ratio per standard deviation (95% confidence interval)p-valueAI-GUR1.60 (1.23, 2.11)0.0003Years since diagnosis1.09 (0.85, 1.38)0.50 Results of a multivariate logistic regression model predicting Gleason upgrade at first follow-up biopsy. Model includes a random effect for study. After accounting for diagnostic GGG, AI-GUR score remained significantly associated with risk of GGG reclassification (p = 0.02), while years since diagnosis and diagnostic GGG were not statistically significant (p = 0.53 and p = 0.33, respectively; Table 3(A)). This indicates that the AI-based portion of the AI-GUR score contributes significant information about the risk of GGG reclassification above and beyond the information contained in the diagnostic GGG.Table 3.Results of multivariable logistic regression models with AI-GUR, years between diagnosis and first follow-up biopsy, and clinical variables predicting Gleason upgrade. Models include a random effect for study.A: AI-GUR, years between diagnosis and first follow-up biopsy, and diagnostic Gleason grade group predicting Gleason upgrade.PredictorOdds Ratio per standard deviation (95% confidence interval)p-valueAI-GUR1.46 (1.06, 2.03)0.02Years since diagnosis1.08 (0.85, 1.37)0.53Diagnostic Gleason Grade Group 0.33 GGG1Reference GGG20.61 (0.21, 1.62) B: AI-GUR, years between diagnosis and first follow-up biopsy, and log(diagnostic PSA + 1) predicting Gleason upgrade.PredictorOdds Ratio per standard deviation (95% confidence interval)p-valueAI-GUR1.60 (1.23, 2.11)3.2 × 10−4Years since diagnosis1.12 (0.88, 1.43)0.35Log(diagnostic PSA + 1)1.30 (1.01, 1.68)0.04 Table 3.Results of multivariable logistic regression models with AI-GUR, years between diagnosis and first follow-up biopsy, and clinical variables predicting Gleason upgrade. Models include a random effect for study.A: AI-GUR, years between diagnosis and first follow-up biopsy, and diagnostic Gleason grade group predicting Gleason upgrade.PredictorOdds Ratio per standard deviation (95% confidence interval)p-valueAI-GUR1.46 (1.06, 2.03)0.02Years since diagnosis1.08 (0.85, 1.37)0.53Diagnostic Gleason Grade Group 0.33 GGG1Reference GGG20.61 (0.21, 1.62) B: AI-GUR, years between diagnosis and first follow-up biopsy, and log(diagnostic PSA + 1) predicting Gleason upgrade.PredictorOdds Ratio per standard deviation (95% confidence interval)p-valueAI-GUR1.60 (1.23, 2.11)3.2 × 10−4Years since diagnosis1.12 (0.88, 1.43)0.35Log(diagnostic PSA + 1)1.30 (1.01, 1.68)0.04 Results of multivariable logistic regression models with AI-GUR, years between diagnosis and first follow-up biopsy, and clinical variables predicting Gleason upgrade. Models include a random effect for study. In exploratory analyses, AI-GUR also added significant predictive information to other clinical variables, including CAPRA score, NCCN risk category, and their components (all p-values < 0.01, not shown). The only clinical feature that added significant predictive information to AI-GUR was diagnostic PSA (transformed as log(PSA +1)), which was only slightly significant (p = 0.04, Table 3(B)). This indicates that the AI-GUR score provides significant, independent information about risk of Gleason upgrade above and beyond the information contained in clinical variables. Additionally, no meaningful correlation was observed between AI-GUR and the CCR Score (CCR) (r = −0.18) or its components (CCP and CAPRA) (r = −0.03 and r = −0.22, respectively) (Figure 2(A-C)), indicating the AI-GUR score contains information independent of these variables. Figure 2.Distributions of AI-GUR compared to (A) the combined clinical risk (CCR) score or its components, (B) the cell-cycle progression (CCP) score and (C) the University of California San Francisco Cancer of the Prostate Risk Assessment (CAPRA) score.The first scatter plot compares AI-GUR Score to CCR Score, showing a weak negative correlation of negative 0.18. The horizontal axis represents CCR Score from negative 1.5 to 1.0 and the vertical axis represents AI-GUR Score from negative 3 to 1. Data points are scattered with no clear clustering, indicating little relationship. The second scatter plot compares AI-GUR Score to CCP Score, with a correlation of negative 0.03. The horizontal axis ranges from negative 3 to 3 and the vertical axis from negative 3 to 1. The distribution is dense in the center, showing minimal correlation. The third scatter plot compares AI-GUR Score to CAPRA Score, with a correlation of negative 0.22. The horizontal axis ranges from 0 to 4 and the vertical axis from negative 3 to 1. Points are aligned vertically, indicating no strong relationship. The legend indicates gray points for 'No upgrade' and green points for 'Upgrade'.Scatter plots showing AI-GUR Score compared to CCR, CCP and CAPRA scores with insigificant negative correlations. Distributions of AI-GUR compared to (A) the combined clinical risk (CCR) score or its components, (B) the cell-cycle progression (CCP) score and (C) the University of California San Francisco Cancer of the Prostate Risk Assessment (CAPRA) score. As part of the PATHOMIQ_PRAD algorithm, cribriform tissue can be reported and analyzed as a binary (present yes or no), as a proportion of total tissue area, or as a proportion of area of cancer tissue. In the validation cohort, AI-identified GP4 cribriform morphology was present in 22.8% (45 of 197) of patients who did not have a GGG reclassification event and in 27.3% (27 of 99) patients with a GGG reclassification event (p = 0.49). Where GP4 cribriform was identified, the extent of cribriform ranged from 0.01% to 37.9% of total AI-identified tissue area, and 0.05% to 60.7% of total AI-identified cancer tissue area. The distribution of extent of cribriform morphology by GGG reclassification status is shown in Supplemental Figure S3. As shown in Table 4(A-C), the presence (binary yes/no) or extent of GP4 cribriform (as proportion of total tissue area or tumor tissue area) did not add significant information to the AI-GUR score when predicting GGG upgrade at confirmatory biopsy (all p-values > 0.1), whereas AI-GUR score remained significant (all p-values < 0.001). These data indicate that the AI-GUR score, which contains information about cribriform morphology, contains more predictive information than does cribriform morphology alone.Table 4.Results of multivariate logistic regression models with AIGUR, years between diagnosis and first follow-up biopsy, and AI-identified cribriform variables predicting Gleason upgrade at first follow-up biopsy. Models include a random effect for study.A: AI-GUR, years between diagnosis and first follow-up biopsy, and binary AI-identified cribriform presence predicting Gleason upgrade.PredictorOdds Ratio per standard deviation (95% confidence interval)p-valueAI-GUR1.60 (1.22, 2.13)0.0005Years since diagnosis1.09 (1.85, 1.38)0.50AI-identified Gleason pattern 4 cribriform 0.95 AbsentReference Present0.98 (0.54, 1.76) B: AI-GUR, years between diagnosis and first follow-up biopsy, and AI-identified proportion of cribriform morphology in identified tissue predicting Gleason upgrade.PredictorOdds Ratio per standard deviation (95% confidence interval)p-valueAI-GUR1.64 (1.24, 2.19)0.0004Years since diagnosis1.09 (0.85, 1.39)0.48Percent of tissue identified as Gleason pattern 4 cribriform0.94 (0.72, 1.23)0.61C: AI-GUR, years between diagnosis and first follow-up biopsy, and AI-identified proportion of cribriform morphology in identified tumor predicting Gleason upgrade.PredictorOdds Ratio per standard deviation (95% confidence interval)p-valueAI-GUR1.71 (1.30, 2.31)9.8 × 10−5Years since diagnosis1.10 (0.86, 1.39)0.45Percent of tumor identified as Gleason pattern 4 cribriform0.81 (0.60, 1.05)0.12 Table 4.Results of multivariate logistic regression models with AIGUR, years between diagnosis and first follow-up biopsy, and AI-identified cribriform variables predicting Gleason upgrade at first follow-up biopsy. Models include a random effect for study.A: AI-GUR, years between diagnosis and first follow-up biopsy, and binary AI-identified cribriform presence predicting Gleason upgrade.PredictorOdds Ratio per standard deviation (95% confidence interval)p-valueAI-GUR1.60 (1.22, 2.13)0.0005Years since diagnosis1.09 (1.85, 1.38)0.50AI-identified Gleason pattern 4 cribriform 0.95 AbsentReference Present0.98 (0.54, 1.76) B: AI-GUR, years between diagnosis and first follow-up biopsy, and AI-identified proportion of cribriform morphology in identified tissue predicting Gleason upgrade.PredictorOdds Ratio per standard deviation (95% confidence interval)p-valueAI-GUR1.64 (1.24, 2.19)0.0004Years since diagnosis1.09 (0.85, 1.39)0.48Percent of tissue identified as Gleason pattern 4 cribriform0.94 (0.72, 1.23)0.61C: AI-GUR, years between diagnosis and first follow-up biopsy, and AI-identified proportion of cribriform morphology in identified tumor predicting Gleason upgrade.PredictorOdds Ratio per standard deviation (95% confidence interval)p-valueAI-GUR1.71 (1.30, 2.31)9.8 × 10−5Years since diagnosis1.10 (0.86, 1.39)0.45Percent of tumor identified as Gleason pattern 4 cribriform0.81 (0.60, 1.05)0.12 Results of multivariate logistic regression models with AIGUR, years between diagnosis and first follow-up biopsy, and AI-identified cribriform variables predicting Gleason upgrade at first follow-up biopsy. Models include a random effect for study. Distributions of AI-GUR The distributions of AI-GUR scores in N = 916 GGG1 patients and N = 478 GGG2 patients who received a CCR score below the AS threshold through routine clinical testing are shown in Figure 3. The distributions differed by diagnostic GGG, indicating that GGG2 patients were less likely to experience GGG reclassification than GGG1 patients. In GGG1 patients, the 20th and 80th percentiles of AI-GUR scores were −1.646 and −0.767, respectively, corresponding to risks of GGG reclassification of 29.64% and 41.75% based on data observed in the validation cohort. The 20th and 80th percentiles of AI-GUR scores in GGG2 patients were −2.885 and −2.151, respectively, corresponding to risks of 16.62% and 23.70%. Figure 3.AI-GUR scores in the distribution analysis set, separated by diagnostic Gleason grade. (A). Diagnostic Gleason 3 + 3, N = 916. Median = −1.181, interquartile range −1.532 to −0.857. (B). Diagnostic Gleason 3 + 4, N = 477. Median = −2.496, interquartile range −2.794 to −2.217.Image A shows a histogram labeled (a) with AI-GUR Scores from -4 to 3 and Counts from 0 to 160. Bars range from -3.1 to 1.0, peaking at -1.2 with a count of 150. Most bars cluster between -2.2 and -0.5, rising to a peak at -1.2, then decreasing. A small right-tail extends from 0 to 1.0 with counts up to 10. Image B shows a histogram labeled (b) with AI-GUR Scores from -4 to 3 and Counts from 0 to 160. Bars range from -3.9 to -0.8, peaking at -2.4 with a count of 90. Most bars cluster between -3.2 and -1.8, increasing to a peak at -2.4, then decreasing. A small right-tail bar appears near -0.5 with counts up to 5.A histogram pair showing AI-GUR score distributions by diagnostic Gleason 3 plus 3 and Gleason 3 plus 4. AI-GUR scores in the distribution analysis set, separated by diagnostic Gleason grade. (A). Diagnostic Gleason 3 + 3, N = 916. Median = −1.181, interquartile range −1.532 to −0.857. (B). Diagnostic Gleason 3 + 4, N = 477. Median = −2.496, interquartile range −2.794 to −2.217. Discussion We developed and validated a novel diagnostic biopsy DPAI model that predicts GGG reclassification events on confirmatory biopsies for patients who are candidates for AS. This model was trained on diagnostic biopsy tissues from patients who were identified as candidates for AS by conventional risk stratification. Validation was conducted in biopsy tissues from patients who were identified as candidates for AS by conventional risk stratification as well as by molecular testing (CCR). Importantly, AI-GUR may address an unmet need described in the NCCN Guidelines, i.e., that there is not an advanced risk stratification tool with sufficient evidence to identify patients at elevated risk of GGG reclassification during the AS candidacy confirmation process [2]. To our knowledge, the AI-GUR model described herein is the first AI-powered H&E image-based morphometric DPAI-derived risk score with validity at this critical decision point. DPAI-based prognostic and predictive models must demonstrate added value beyond conventional pathological features routinely used in clinical decision-making, including GGG and cribriform morphology. MRI data were not available for the cohorts used in this study. However, the clinical relevance of MRI-derived quantitative imaging models (including radiomics) is underscored by the incorporation of mpMRI into contemporary diagnostic pathways and by risk calculators that include MRI factors that are in use to predict clinically significant disease [9]. In the present study, AI-GUR remained independently predictive of GGG reclassification after accounting for diagnostic GGG and AI-identified cribriform morphology. Additionally, AI-GUR provided information that was not strongly correlated with prognostic risk methodologies (CAPRA, Prolaris). AI-GUR represents a novel enhancement of a genomic classifier that provides individual risk related to the critical oncological outcomes of distant metastasis and PCa-specific mortality. The model was architected to be complementary to the existing CCR prostate cancer prognostic framework, and the validation cohort was identified by a genomic classifier result that indicated candidacy for AS. This result is based on having a <3.2% risk of distant metastasis 10 years after diagnosis if disease is managed by non-interventional means (e.g., AS). Importantly, in validation studies of the genomic classifier, no patient identified as a candidate for AS succumbed to their PCa [3]. Thus, AI-GUR has potential utility above and beyond what is currently available as a decision aid for AS candidates and their health care providers. In this study, a small number of NCCN unfavorable intermediate-risk patients were considered eligible for AS based on a combination of clinical and molecular features. Given the clinical features of their PCa, patients with unfavorable intermediate risk disease are not typically considered AS candidates and are not routinely offered AS as a disease management strategy by their health care providers; however, a small portion of these patients may have more favorable prognoses when also considering molecular features. As such, these patients were not excluded from analysis. Whereas standard clinical risk variables such as GGG and NCCN risk group did not provide significant information to AI-GUR, PSA at diagnosis provided marginally significant (p = 0.04) information. This relationship should be further explored, especially given the weight PSA has in the CANARY PASS Calculator [10]. Cribriform morphology is definitionally absent from GGG1 and increases in prevalence as GGG increases. In the ProtecT study [11], cribriform morphology was identified in ~13% of the pathologically re-reviewed cohort, and evidence of cribriform positivity was associated with significantly higher risk of developing metastases [12]. The prior literature is mixed related to whether extent of cribriform morphology has differential impact on outcomes, some studies suggesting large cribriform architecture being associated with particularly aggressive disease [13,14], whereas a biopsy-based study with long-term follow-up indicated that both large and small cribriform morphology were similarly associated with poor outcomes [15]. As demonstrated herein, the AI-GUR predictive model differentiates risk of GGG reclassification events independent of the cribriform sub-feature of the PATHOMIQ_PRAD algorithm. While AI-GUR predicts a higher Gleason score on subsequent biopsy, AI-GUR was not architected as prescriptive, and moving from GGG1 to GGG2 does not necessitate transition from AS to definitive therapy. Further validation in independent cohorts will determine the impact of biopsy modality (systematic, image-guided) on the performance of AI-GUR. While these studies are underway, this is a limitation of the current study. If AI-GUR was predicting true disease progression, the distribution of upgrade events would be enriched at later time points, which was not observed in first follow-up biopsies as far as 5 years post-diagnosis (Supplemental Figure S2). Although the model does not directly distinguish disease progression from biopsy-sampling variability (due to factors such as tumor heterogeneity) that distinction would not impact clinical decisions as currently understood: a confirmatory biopsy has importance irrespective of the cause of reclassification. Another potential limitation is selection bias, as included patients were already AS candidates who underwent genomic classifier testing. However, limiting inclusion to AS candidates was intentional to the study design, given that the objective was to develop a tool specific to this population. The inclusion of patients ineligible for AS could have confounded the results, resulting in a model not relevant to the intended use case. DPAI-enabled models, such as AI-GUR, have promising potential as decision aids in the management of PCa, and future work will evaluate its utility in real-world scenarios. Conclusions In this study, we report the development and validation of a DPAI-derived risk score that predicts Gleason upgrade on subsequent biopsy for patients that are candidates for AS. This new metric provides information beyond conventional risk stratification tools and provides individualized risk estimates that may improve shared decision-making for AS candidates. Supplementary Material Suppl Figure 2_300dpi.png Suppl Figure 3B_300dpi.png TRIPOD AI Checklist_22MAY2026.pdf Suppl Figure 1_300dpi.png Supplemental Table_20MAY2026.docx Suppl Figure 3A_300dpi.png