Abstract
Background and Objectives
Biochemical recurrence (BCR) after radical prostatectomy (RP) is a heterogeneous disease state in prostate cancer with multiple treatment options. Improved risk stratification could enable more personalized decision-making. We developed and validated a digital pathology-based multi-modal artificial intelligence (MMAI) model to predict outcomes in post-RP BCR patients undergoing salvage therapy.
Methods
An MMAI model was trained to predict distant metastasis (DM) using prostate histopathology image features and clinical variables (pathologic Grade Group, pathologic T-stage, pre-salvage radiotherapy (SRT) PSA, age, and surgical margin). The locked model was validated in 533 patients from NRG/RTOG 9601 and 0534 treated with SRT +/− hormone therapy (HT), using Cox regression and time-dependent AUC.
Key Findings and Limitations
With a median follow-up of 9.3 years, MMAI score was significantly associated with DM (subdistribution hazard ratio = 2.17 per SD (95% CI 1.65-2.85; p<0.001) and remained independently prognostic after adjusting for clinical variables and treatment. The 10-year tdAUC for MMAI was 0.74 compared to 0.68 for a clinical nomogram. Binary risk categorization demonstrated higher 10-year DM incidence in MMAI High Risk (25%) vs. Low Risk (8.8%). The absolute reduction in 10-year DM with HT plus SRT vs. SRT alone was 21% in High Risk vs. 2.5% in Low Risk. Limitations include use of archived trial cohorts.
Conclusions and Clinical Implications
The Post-RP MMAI model provides individualized risk estimates after SRT +/− HT and may support shared decision-making about salvage treatment. External and prospective validation are ongoing.
Keywords: Artificial intelligence, digital pathology, prostate cancer, radical prostatectomy, salvage therapy, prognostic model
Introduction
After radical prostatectomy (RP) for localized prostate cancer, approximately 30-85% of men will require subsequent therapy for biochemical recurrence (BCR) [1]. Randomized trials have evaluated post-RP radiotherapy to reduce the risk of distant metastasis (DM), with many patients rendered disease-free long-term. Given the small but significantly higher toxicity rates from adjuvant radiotherapy (ART), early salvage radiotherapy (SRT) is now favored, with recent data showing comparable outcomes to ART [2]. However, outcomes post-SRT remain heterogeneous [3], and with results from the RADICALS-HD trial failing to show a clear benefit of adding hormone therapy (HT) to SRT for most post-RP patients [4,5], there is clear need for improved risk stratification in this setting.
Existing prognostic tools have been developed and validated mainly in localized and post-RP populations, with limited validation in post-RP BCR settings. Consequently, most phase III trials of post-RP SRT±HT enrolled relatively unselected patients, contributing to modest observed HT survival benefit, since many had indolent disease manageable with SRT alone. This has led to both under- and over-treatment of men with BCR post-RP, highlighting an unmet need for more individualized treatment approaches [2]. Moreover, outcome disparities across sociodemographic groups [6] highlight the need for accessible biomarkers that capture biological heterogeneity.
Recent advances in digital pathology and artificial intelligence (AI) have enabled novel prognostic and predictive biomarkers from prostate biopsy tissue [7]. These multimodal AI (MMAI) algorithms outperform standard NCCN risk groups and predict HT benefit in newly diagnosed disease [7,8], with validation across racial subgroups and diverse clinical settings [9,10]. However, these tools were not intended for use in the post-RP BCR setting. Therefore, we trained and validated a novel prognostic biomarker for patients with BCR post-RP using multimodal deep learning on digital histopathology images and clinical data from two phase III trials of SRT±HT.
Materials and Methods
Cohort Details
Patients were selected from two NRG/RTOG Phase 3 trials, 9601 [11] and 0534 [12] (NCT00002874, NCT00567580), which enrolled men with BCR post-RP receiving SRT±HT between 1998-2015. Both trials centrally-collected RP histopathology slides and reported long-term outcomes. In NRG/RTOG 9601, patients were randomized to SRT±long-term bicalutamide HT. In NRG/RTOG 0534, patients were randomized to salvage prostate bed RT (PBRT) ± short-term (ST)-ADT ± pelvic lymph node RT. Trial details for NRG/RTOG 9601 and 0534 have been described previously [11,12]IRB approval was obtained from NRG Oncology (IRB00000781) and informed consent waived because this study used anonymized data.
Patients were eligible for inclusion if digitized H&E-stained histopathology slides, baseline clinical variables, and DM outcomes were available. Image acquisition details are outlined in the Supplementary Appendix. NRG/RTOG 9601 and 0534 patients were randomly split 70%/30% into a development and held-out validation set, stratified by trial, DM status, and pathologic Gleason Grade Group (GG). This split was pre-specified to preserve an untouched internal validation set not accessible to the AI development team. An additional histopathology image set spanning multiple cancers was available for image model development (Supplementary Appendix).
Objective and Endpoints
The primary objective of this translational post-hoc study was to develop and validate an AI-based post-RP BCR model to stratify DM risk. The primary endpoint was DM (from time of randomization). Secondary endpoints included time from randomization to death with DM (DDM; death of any cause with a DM record), biochemical failure (BF; defined by trial protocols and/or Phoenix definition), freedom from progression (FFP; progression defined as BF, clinical failure (local, regional, or distant), or death from any cause), metastasis-free survival (MFS), prostate cancer-specific mortality (PCSM defined by trial protocols), and overall survival (OS). Patients without events were censored at last follow-up.
Model Development
The model development process is depicted in Supplementary Figure 1 and consisted of image preprocessing to divide digital slide images into patches and remove patches without usable tissue (Supplementary Appendix), self-supervised learning (SSL) image feature extraction, and a multimodal fusion pipeline to associate image and pre-specified prognostic clinical features with clinical outcomes.
The multimodal model architecture consisted of: (1) a clinical Cox proportional hazards model trained on age, pre-SRT PSA, pathologic T-stage, pathologic Gleason GG, and surgical margin status, (2) an image model to learn the association between SSL image features and clinical outcomes, and (3) a late fusion pipeline to aggregate the predictions from image and clinical models. The image model was developed with prostate histopathology image data using an attention-based multiple instance learning framework [13]. Once downstream clinical and image models were trained, their parameters were frozen prior to fusion, which standardized the output predictions of each model relative to their respective distributions in the training data and then averaged the predictions to generate a continuous relative risk score from 0 to 1. After training, the Post-RP MMAI (v1.1) was locked and evaluated in the validation cohort.
Statistical Analysis
Validation of the locked model was conducted using the held-out 30% of patients from NRG/RTOG 9601 (N=185) and 0534 (N=359). A pre-planned power calculation indicated 80% power to detect a hazard ratio (HR) of 1.38 per standard deviation increase in MMAI score, assuming proportional hazards and a trial-based 14.4% DM rate.
Prognostication was assessed using univariable Fine and Gray (F&G) regression [14] for endpoints with competing risks (DM, DDM, PCSM, BF, FFP), treating death without the event as a competing risk, and Cox regression for MFS and OS. Results were reported as subdistribution HR (sHR) or HR with 95% confidence interval (CI). Schoenfeld residuals for DM showed no violation of the proportional hazard assumption. DM was further assessed using the time-dependent area under the ROC curve (tdAUC) [14,15] at 10 years, accounting for competing events, with 95% CIs from 1000 bootstrap replicates. Multiplicity adjustments were not applied to secondary endpoints. Because the Post-RP MMAI model outputs a continuous risk score, rather than absolute predicted probabilities, standard calibration plots were not used. Decision curve analysis (DCA) was performed for censored outcomes with competing risk [16].
Binary risk groups were derived in the development test split by evaluating a range of MMAI score cut-points and selecting the 60th percentile based on separation of cumulative incidence functions [17] for DM and PCSM. This percentile-based threshold was identified for statistical interpretability, locked and subsequently applied unchanged to the validation set, where separation of outcome risk was assessed using 5- and 10-year estimates, 95% CIs, and Gray’s test p-values.
Multivariable F&G regression adjusted for all clinical input variables was performed to evaluate the added value of the MMAI over a clinical variable-only nomogram for SRT after RP [18], as well as for treatment type to account for therapies received. To examine predictive potential, a multivariable F&G regression with biomarker-treatment interaction was performed; the Wald test assessed interaction significance. This study adheres to TRIPOD-AI statement guidelines (Supplementary Appendix).
Descriptive statistics were reported as counts and proportion (%) for categorical variables, or median and interquartile range (IQR) for continuous variables. All tests were two-sided with ɑ=0.05.
Software
Statistical analyses were performed using R (v4.4.0, RFoundation for Statistical Computing, Vienna, Austria). The MMAI model was implemented in Python (v3.10.17, Python Software Foundation, https://www.python.org/) and PyTorch v2.7.0 [19].
Results
Cohort Characteristics
Of the 2632 patients enrolled in NRG/RTOG 0534 and 9601, 1921 of the 2573 (75%) patients with available clinical data had histopathology slides at the NRG Biospecimen Bank. Of these patients, 1855 (97%) had RP specimen images and clinical data, with 1,322 and 533 patients used in the development and validation sets, respectively (Figure 1). Baseline characteristics are shown for patients in these development and validation sets (Table 1), as well as for patients in the validation set separated by MMAI risk groups and treatment groups (Supplementary Table 1, Supplementary Table 2). The median number of slides per patient was 2 (range 1-62). The validation cohort had a median follow-up of 9.2 years for censored patients (IQR: 7.1-11.3), the median age was 64 years (IQR: 59-69), median pre-SRT PSA was 0.4 ng/mL (IQR: 0.3-0.8), and 9.2% patients were Black. The majority (65%) of patients had pathologic GG2 or 3. MMAI score statistics in the validation cohort are shown in Supplementary Table 3.
Figure 1.
Patient flow diagram for development and validation cohort.
Table 1.
Patient characteristics for overall, development, and validation cohorts
| Cohort (N=1,855) | ||
|---|---|---|
| Variable | Developmenta N = 1,322 |
Validationa N = 533 |
| Treatment arm | ||
| RT + bicalutamide | 224 (17%) | 94 (18%) |
| RT + STADT | 575 (43%) | 247 (46%) |
| RT only | 523 (40%) | 192 (36%) |
| Age (years) | 64 (59, 69) | 64 (59, 69) |
| Race | ||
| African American | 133 (10%) | 49 (9.2%) |
| Other | 30 (2.3%) | 9 (1.7%) |
| Unknown | 20 (1.5%) | 9 (1.7%) |
| White | 1,135 (86%) | 466 (87%) |
| (Missing) | 4 | 0 |
| Pre-salvage RT PSA (ng/mL) | 0.4 (0.3, 0.8) | 0.4 (0.3, 0.8) |
| Pathological T Stage | ||
| T2 | 490 (37%) | 175 (33%) |
| T2a | 23 (1.7%) | 12 (2.3%) |
| T2b | 104 (7.9%) | 42 (7.9%) |
| T3a | 480 (36%) | 115 (22%) |
| T3b | 225 (17%) | 89 (17%) |
| Pathological N Stage | ||
| N0 | 1,278 (100%) | 514 (100%) |
| (Missing) | 44 | 19 |
| Pathological Grade Group | ||
| 1 | 252 (19%) | 97 (18%) |
| 2 | 521 (40%) | 216 (41%) |
| 3 | 310 (24%) | 127 (24%) |
| 4 | 127 (9.7%) | 53 (9.9%) |
| 5 | 90 (6.9%) | 40 (7.5%) |
| (Missing) | 22 | 0 |
| Median follow-up for censored patients (years) | 9.3 (7.3, 12) | 9.2 (7.1, 11) |
Data are reported as median (IQR) or frequency (%)
Abbreviations: RT, radiation therapy; PSA, prostate-specific antigen; STADT, short-term androgen deprivation therapy
MMAI Model Validation
The Post-RP MMAI score was prognostic across all endpoints in the validation cohort (Figure 2), demonstrating consistent associations across multiple clinically relevant outcomes, including DM, DDM, PCSM, BF, FFP, MFS, and OS. For the primary endpoint of DM, the sHR per SD increase of MMAI score was 2.17 (95%CI: 1.65-2.85, p<0.001). The association remained significant after adjusting for all clinical input variables and treatment type (sHR 2.56, 95%CI 1.54-4.25) (Supplementary Table 4), as well as after adjusting for a clinical nomogram [18] (Supplementary Table 5). The 10-year DM tdAUC of the MMAI score was 0.74 (95%CI: 0.65-0.81), compared with 0.68 (95%CI: 0.59-0.75) for the clinical nomogram, and 0.70 (95%CI: 0.62-0.78) for the Cox model containing all clinical input variables. DCA for the 10-year DM risk further demonstrated that the MMAI model provided greater net benefit across a broad range of threshold probabilities compared with the clinical nomogram and clinical Cox model (Supplementary Figure 2).
Figure 2.
Prognostic associations of the Post-RP MMAI across clinical endpoints in the validation cohort. Subdistribution hazard ratios (sHR) or hazard ratios (HR) per 1 standard deviation increase in MMAI score are reported with 95% confidence intervals (CI) across each endpoint.
Cumulative incidence analysis revealed a higher 5-year DM incidence among the MMAI High-Risk group (9.8%, 95%CI: 6.4-14%) compared to Low-Risk (4.4%, 95%CI: 2.5-7.2%), respectively; Figure 3); at 10-year, these were 25% (95%CI: 19-32%) versus 8.8% (95%CI: 5.7-13%), respectively. For PCSM, the 5-year cumulative incidence was 3.2% (95%CI: 1.4-6.1%) in the High-Risk group versus 0.3% (95%CI: 0.0-1.8%) in the Low-Risk group; the 10-year PCSM rates were 139% (95%CI: 8.3-18%) versus 3.9% (95%CI: 1.9-7.1%).
Figure 3.
Cumulative incidence curves for A) distant metastasis (DM) and B) prostate cancer-specific mortality (PCSM) in the validation cohort. Shaded areas indicate the 95% confidence intervals. Gray’s test p-value was <0.001 for the primary endpoint (DM).
The 10-year DM rates between patients receiving SRT-only versus SRT+HT were 38% (95%CI: 26-50%) versus 17% (95%CI: 11-25%) among the MMAI High-Risk group (HR 2.10, 95%CI 1.22-3.61) and 10% (95%CI: 4.9-18%) versus 7.8% (95%CI: 4.4-12%) among the MMAI Low-Risk group (HR 1.27, 95%CI 0.59-2.72) (Figure 4). Despite the narrower effect size in Low-Risk patients, a statistically significant biomarker-treatment interaction effect was not observed (p-int=0.3).
Figure 4.
Hormone therapy (HT) benefit among RP MMAI risk groups: Cumulative incidence curves for distant metastasis (DM) within RP MMAI High Risk group (left) and RP MMAI Low Risk group (right). Subdistribution hazard ratios (sHR) are reported with 95% confidence interval (CI) and Wald test p-value from Fine-Gray proportional hazard models are shown.
To enhance interpretability in the model’s predictions, we reviewed image patches with the highest and lowest attention assigned by the image model. Pathologist evaluation of these image patches supported the model’s prioritization of biologically and clinically relevant features, such as cribriform and solid morphology, when generating patient-level MMAI risk predictions (Supplementary Figure 3).
Discussion
In this study, we developed and validated a novel Post-RP MMAI model combining deep learning-derived histopathology features and routine clinical variables to predict outcomes in men with BCR following RP. The Post-RP MMAI model is the first biomarker that, to our knowledge, has been trained and validated using large, phase III trials in this setting. This captures heterogeneity across hundreds of centers across North America, spanning academic, community, and Veteran Affairs practices, with variability in tissue age/processing, and patient-specific factors. Together with previous analytic validation work for MMAI models [20], these features support enhanced accuracy, robustness, and generalizability for clinical use. The MMAI approach is also feasible to implement in laboratories with digital pathology infrastructure, and future cloud-based deployment could further broaden access.
The Post-RP MMAI model showed robust performance across all evaluated outcomes, and identified differential benefit from the addition of HT to SRT. In MMAI High-Risk patients, the absolute 10-year DM reduction with SRT+HT was 21% (number needed to treat (NNT) 5), vs. 2.5% (NNT 40) in MMAI Low-Risk. These findings suggest that the Post-RP MMAI risk score could help personalize HT use in the post-RP salvage setting. While PSA kinetic information commonly used to guide prognostication following post-RP BCR (e.g. PSA doubling time) was not available, these results support MMAI’s potential to guide treatment intensification or de-intensification in this population. More generally, this work builds on a growing field of pathology AI in prostate cancer, which has largely focused on diagnostic applications such as Gleason grading [21]. While other prognostic models remain earlier in development with limited randomized validation [22,23], the pathology AI approach holds promise both in terms of model performance and in potentially streamlined workflow. There remains substantial opportunity for development of additional pathology-based AI models in prostate cancer, with the quality of the development and validation datasets likely being critical drivers of model performance in the clinical setting.
In MMAI modelling, inclusion of established predictors such as PSA and GG can obscure whether there is meaningful contribution from image features. In our study, the MMAI model remained independently prognostic after adjusting for clinical variables and demonstrated greater net benefit in decision curve analysis. Collectively, these findings indicate that the Post-RP MMAI adds prognostic value beyond clinicopathologic variables. Although exploratory visualization of high-attention regions aligned with known adverse morphology such as cribriform patterns (Supplementary Figure 3), studies are ongoing to improve interpretability and link image features to tumor biology for enhanced explainability.
Ongoing validation work is needed to ensure reproducibility and generalizability in independent datasets. Indeed, while the current analyses constitute internal validation, external validation has since shown prognostic performance of the Post-RP MMAI model in an independent institutional post-surgical cohort [24]. Nonetheless, further validation in underrepresented subgroups, such as racial and ethnic minority groups, will be critical. An additional limitation of the study is that the model did not explicitly account for the use of pelvic nodal RT or HT variability. While this was done to help ensure prognostic performance irrespective of treatment, absolute risk estimates will vary based on the exact treatment received. Moreover, because of the trial enrollment era, PSMA PET/CT was not routinely used, potentially underestimating disease burden. Additionally, modern therapeutic options were not yet available, which may affect how risk stratification translates to current practice. As MMAI adoption progresses, future work should evaluate its cost-effectiveness and decisional utility in real-world implementation.
The MMAI model outputs a continuous relative risk score rather than absolute predicted probabilities, which limits formal calibration assessment; instead, we evaluated risk stratification through cumulative incidence curves and event rates across risk groups. Nonetheless, DCA results should be interpreted in that context, recognizing that comparisons may be influenced by differences in calibration, particularly between the prespecified clinical nomogram, which can be miscalibrated, and the clinical Cox and MMAI models, for which absolute risk estimates used in the DCA are derived from the underlying trial data. Nonetheless, the model’s apparent benefit above a 10% threshold, which is specific to this dataset and may translate differently in contemporary datasets, aligns with the baseline 10-year metastatic risk reported in GETUG-AFU 16 [25], RTOG 9601 [11], RTOG 0534 [12] where ADT provided meaningful benefit, indicating that this range corresponds to the level of metastatic risk where treatment-intensification decisions are typically considered in practice. Beyond identifying patients most likely to benefit from ADT with salvage radiotherapy, the MMAI could also inform related treatment-intensification decisions, which warrant evaluation in future trial datasets.
Conclusions
The Post-RP MMAI model provides a promising tool to estimate DM risk after SRT±HT for men with BCR post-RP. This approach can support shared decision-making for personalizing HT use. To further validate and expand its clinical utility, the Post-RP MMAI model will be operationalized as a laboratory developed test and is currently being prospectively evaluated in the PROSTATE-IQ trial (NCT06274047) to assess clinical actionability.
Supplementary Material
Advancing Practice:
What does the study add?
A novel digital pathology-based multimodal artificial intelligence (MMAI) biomarker was developed to predict long-term outcomes in men with biochemical recurrence (BCR) following radical prostatectomy (RP). The model was trained using digitized histopathology images and clinical variables from prospective phase III clinical trials and validated in a cohort of post-RP BCR patients from NRG/RTOG 9601 and 0534. The MMAI model was independently associated with distance metastasis after adjustment for clinical factors and stratified patients into risk groups that may garner differential benefit from the addition of hormone therapy. These findings support the post-RP BCR MMAI model as a potential tool to improve risk stratification and individualize treatment decisions in men receiving salvage radiotherapy following prostatectomy.
Patient Summary:
In this study, we developed and tested an artificial intelligence (AI) tool that analyzes prostate tissue images and clinical information to predict long-term outcomes after prostate surgery. This AI tool may help doctors and patients make more informed decisions about additional treatments if PSA levels rise, supporting more personalized therapy choices.
Acknowledgments:
We thank NRG staff Sheralee Miller, Leslie Longoria, Florence Lo, Jen Chieh-Lee, Michael Yuen, Sam Clinch, and Suzanne Baldwin for their administrative support of the project submission to the Cancer Therapy Evaluation Program (CTEP) National Clinical Trials Network Core Correlative Sciences Committee (NCTN-CCSC) and for the submission of this manuscript.
Funding:
Research reported in this publication was supported by the National Cancer Institute of the National Institutes of Health under Award Number UG1CA189867 (NCORP), U10CA180822 (NRG Oncology SDMC), U10CA180868 (NRG Oncology Operations), U24CA196067 (NRG Specimen Bank), and R01CA240991 (Todd M. Morgan and Daniel E. Spratt). The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.
This project was also supported by Artera Inc.
Conflicts of Interest
Drs Bahary, Balogh Greenberg, Hoffman, Ross, and Wilke have nothing to disclose.
Dr Chen declares in the last 36 months support for attending meetings and/or travel from Artera; stock or stock options received from Artera; other financial or non-financial interests from Employee at Artera.
Ms DeVries declares since the initial planning of the work all support for the present manuscript from NCI NCTN NRG Oncology Biospecimen U24 Grant.
Dr Efstathiou declares in the last 36 months consulting fees received from Blue Earth Diagnostics, Boston Scientific, AstraZeneca, Genentech, Clarity Pharmaceuticals; payment or honoraria for lectures, presentations, speakers, bureaus, manuscript writing or educational events from Elekta, IBA, UpToDate; participation on a Data Safety Monitoring Board or Advisory Board for Merck, Roivant Pharma, Myovant Sciences, Janssen, Bayer Healthcare, Progenics Pharmaceuticals, Pfizer, Astellas, Gilead, Lantheus, Blue Earth Diagnostics, Angiodynamics; leadership or fiduciary role in other board, society, committee, or advocacy group, paid or unpaid Board member (unpaid): Massachusetts Prostate Cancer Coalition (MPCC); American College of Radiation Oncology (ACRO); Radiation Oncology Institute (ROI).
Dr Esteva declares since the initial planning of the work all support for the present manuscript from ArteraAI with specifications/comments of I am employed full-time by ArteraAI; support for attending meetings and/or travel from ArteraAI; leadership or fiduciary role in other board, society, committee, or advocacy group, paid or unpaid from ArteraAI with specifications/comments of I am on the board of ArteraAI; stock or stock options received from ArtearAI with specifications/comments of I am a shareholder of ArteraAI.
Dr Feng declares in the last 36 months grants or contracts from Advisor for Artera, Astellas, Bayer, Blue Earth Diagnostics, BMS, ClearNote, Janssen, Myovant, Point Biophar. Research grant from the Prostate Cancer Foundation; Consulting fees received from Astellas, Bayer, Blue Earth Diagnostics, BMS, ClearNote, Janssen, Myovant, Point Biophar; leadership or fiduciary role in other board, society, committee, or advocacy group, paid or unpaid for GU Chair NRG.
Dr Huang declares since the initial planning of the work all support for the present manuscript from Employee of Artera; Stock or stock options received from Employee of Artera.
Dr Martin declares in the last 36 months payment or honoraria for lectures, presentations, speakers, bureaus, manuscript writing or educational events for lecture sponsored by Johnson & Johnson; support for attending meetings and/or travel for travel grant from TerSera Travel Grant from Tolmar.
Dr Michalski declares since the initial planning of the work all support for the present manuscript from NCI with specifications/comments of payments made to institution.
Dr Mitani declares since the initial planning of the work all support for the present manuscript from Artera with specifications/comments of employed; stock or stock options received from Artera.
Dr Morgan declares in the last 36 months grants or contracts from NCI (R01); consulting fees received from Foundation Medicine, Tempus AI with specifications/comments of Advisory boards (personal).
Dr Pollack declares since the initial planning of the work all support for the present manuscript from NIH/NCI P30 CA240139 (PI: Nimer), University of Miami, Sylvester Comprehensive Cancer Center Support Grant paid to institution.
Dr Pugh declares in the last 36 months grants or contracts from NCI with specifications/comments of payments made to institution.
Dr Ren declares since the initial planning of the work all support for the present manuscript from Artera, Inc. with specifications/comments of employee at Artera; support for attending meetings and/or travel from Artera, Inc. with specifications/comments of employee at Artera; stock or stock options received with specifications/comments of employee at Artera.
Dr Royce declares since the initial planning of the work all support for the present manuscript from Artera with specifications/comments of employee of Artera (and equity) at time of writing this manuscript.
Dr Sandler declares since the initial planning of the work all support for the present manuscript from ACR-NRG Oncology with specifications/comments of Chair, GU Cancer Committee of RTOG-NRG Oncology; leadership or fiduciary role in other board, society, committee, or advocacy group, paid or unpaid from ASTRO with specifications/comments of Member, Board of Directors.
Dr Simko declares since the initial planning of the work all support for the present manuscript from National Cancer Institute, USA with specifications/comments of grant support. Dr Spratt declares in the last 36 months consulting fees received from Boston Scientific; payment or honoraria for lectures, presentations, speakers, bureaus, manuscript writing or educational events from Janssen; participation on a Data Safety Monitoring Board or Advisory Board for Astellas, AstraZeneca, Bayer, Pfizer, Novartis; leadership or fiduciary role in other board, society, committee, or advocacy group, paid or unpaid for NCCN Guidelines.
Dr Tang declares since the initial planning of the work all support for the present manuscript from Aretra Inc. with specifications/comments of employed by Artera Inc.; stock or stock options received from Aretra Inc. with specifications/comments of received stock options offered by Artera as an employee.
Dr Tran declares in the last 36 months grants or contracts from NIH, DOD, Prostate Cancer Foundation, Movember Foundation, Distinguished Gentlemen’s Ride Foundation with specifications/comments of institutional grants; royalties or licenses received from Holds a patent 9114158-Compounds and Methods of Use in Ablative Radiotherapy licensed to Natsar Pharm with specifications/comments of royalities; consulting fees received from consultant for Janssen-Taris Biomedical, Bayer Healthcare and RefleXion, personal fees from Janssen-Taris Biomedical, Lantheus, Pfizer, Myovant and AstraZeneca with specifications/comments of self; support for attending meetings and/or travel from Janssen-Taris Biomedical & Bayer Healthcare with specifications/comments of self; leadership or fiduciary role in other board, society, committee, or advocacy group, paid or unpaid from ASTRO and NRG Oncology.
Dr Zwerink declares in the last 36 months grants or contracts from employee at Artera Inc.; stock or stock options received From Artera Inc.
Dr Nguyen declares in the last 36 months grants or contracts from Astellas, Bayer, Janssen with specifications/comments of to my institution for clinical trial support; consulting fees received from Astellas, Bayer, Janssen, Boston Scientific, Nanocan, Blue Earth, Myovant, Novartis, AIQ, Amgen; leadership or fiduciary role in other board, society, committee, or advocacy group, paid or unpaid for NRG Oncology GU Leader, ASTRO Refresher Course Chair; stock or stock options received from Nanocan, Stratagen Bio, Reversal Therapeutics.
Appendix
Dr. Morgan and NRG Oncology would like to thank the following accrual institutions who have been part of our mission to improve the lives of patients with cancer by conducting practice-changing multi-institutional clinical and translational research.
Anne Arundel Medical Center
Arizona Center for Cancer Care-Peoria
Arizona Oncology Services Foundation
Billings Clinic Cancer Center
Boston Medical Center MBCCOP
Brooke Army Medical Center
CarolinaEast Health System-Medical Center
Case Western Reserve University
Centre Hospitalier de lÙniversité de Montréal-Notre Dame
Centre Hospitalier Universitaire de Sherbrooke
Christiana Care Health System-Christiana Hospital
CHUM
Coastal Carolina Radiation Oncology
Columbia St. Mary's Hospital Ozaukee, Inc
Columbus Community Clinical Oncology Program
Cooper Health System-Voorhees
Cross Cancer Institute
Dartmouth Hitchcock Medical Center
David Grant United States Air Force Medical Center
Dubs Cancer Center at Rogue Valley Medical Center
Edward Hines, Jr. VA Hospital
Elkhart General Hospital
Ellis Fischel Cancer Center, University of Missouri
Flower Hospital
Fox Chase Cancer Center
Geisinger Medical Center
Greenville Health System Cancer Institute-Eastside
Henry Ford Hospital
Hunter Holmes McGuire Veterans Administration Medical Center
Huntington Memorial Hospital
Intermountain Medical Center
John H Stroger Jr Hospital of Cook County
Kaiser Permanente Medical Center - Santa Clara
Kalamazoo CCOP-West Michigan Cancer Center
Kansas City CCOP
L Hotel-Dieu de Quebec
LAC + USC Medical Center
Lankenau Medical Center
London Regional Cancer Program
Louisiana State University Health Science Center
Massachusetts General Hospital
Mayo Clinic
McLaren Cancer Institute - Lapeer
Medical University of South Carolina
Memorial Medical Center-Las Cruces
Memorial Sloan Kettering-Basking Ridge
Methodist Medical Center of Illinois
Nevada Cancer Research Foundation CCOP
Northwest Community Clinical Oncology Program
Northwest Community Hospital
Nova Scotia Cancer Centre
Novant Health Presbyterian Medical Center
Ochsner Medical Center Jefferson
Ohio State University Medical Center
ProHealth Oconomowoc Memorial Hospital
Providence Medford Medical Center
Queen's Medical Center
Radiation Medical Group, Inc.
Radiological Associates of Sacramento
Rapid City Regional Hospital
Regional Hospital of Scranton
Regions Hospital
Rex Cancer Center
Richard L. Roudebush VA Medical Center
Roswell Park Cancer Institute
Saint Alphonsus Cancer Care Center-Boise
Saint Mary's Hospital and Regional Medical Center
Saint Vincent Hospital Cancer Center Green Bay
Sarasota Radiation & Medical Oncology Center
Sentara Virginia Beach General Hospital
Sequoia Regional Cancer Center
Shields Radiation Oncology Center
Skagit Valley Hospital
Spartanburg Medical Center
St Luke's-Roosevelt Hospital Center
St. Joseph Hospital
Sutter General Hospital
The Hospital of Central Connecticut
Thomas Jefferson University Hospital
Tom Baker Cancer Centre
Tulane University Medical Center
UC San Diego Moores Cancer Center
UCSF Medical Center-Mount Zion
University Health Network-Princess Margaret Hospital
University of Alabama at Birmingham Medical Center
University of California Davis Medical Center
University of Chicago Comprehensive Cancer Center
University of Florida Health Science Center
University of Michigan Medical Center
University of North Carolina - Chapel Hill
University of Oklahoma Health Sciences Center
University of Utah Health Science Center
UPMC-Magee Women's Hospital
USON-Arizona Radiation Oncology
Veteran Affairs New York Harbor Healthcare System-Brooklyn Campus
Virginia Mason CCOP
Washington University
West Allis Memorial Hospital
William Beaumont Hospital-Royal Oak
Data Sharing Statement:
The data from the present publication will be made available by request from the NCTN/NCORP Data Archive: https://nctn-data-archive.nci.nih.gov. The source code may be made available, subject to intellectual property constraints, by contacting AE (aesteva@artera.ai).
References
- [1].Dess RT, Morgan TM, Nguyen PL, Mehra R, Sandler HM, Feng FY, et al. Adjuvant Versus Early Salvage Radiation Therapy Following Radical Prostatectomy for Men with Localized Prostate Cancer. Curr Urol Rep 2017;18:55. 10.1007/s11934-017-0700-0. [DOI] [PubMed] [Google Scholar]
- [2].Vale CL, Fisher D, Kneebone A, Parker C, Pearse M, Richaud P, et al. Adjuvant or early salvage radiotherapy for the treatment of localised and locally advanced prostate cancer: a prospectively planned systematic review and meta-analysis of aggregate data. The Lancet 2020;396:1422–31. 10.1016/S0140-6736(20)31952-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [3].Morgan TM, Boorjian SA, Buyyounouski MK, Chapin BF, Chen DYT, Cheng HH, et al. Salvage Therapy for Prostate Cancer: AUA/ASTRO/SUO Guideline Part I: Introduction and Treatment Decision-Making at the Time of Suspected Biochemical Recurrence after Radical Prostatectomy. Journal of Urology 2024;211:509–17. 10.1097/JU.0000000000003892. [DOI] [PubMed] [Google Scholar]
- [4].Parker CC, Clarke NW, Cook AD, Petersen PM, Catton CN, Cross WR, et al. Randomised Trial of No, Short-term, or Long-term Androgen Deprivation Therapy with Postoperative Radiotherapy After Radical Prostatectomy: Results from the Three-way Comparison of RADICALS-HD (NCT00541047). European Urology 2024;86:422–30. 10.1016/j.eururo.2024.07.026. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [5].Parker CC, Clarke NW, Cook AD, Kynaston H, Catton CN, Cross WR, et al. Adding 6 months of androgen deprivation therapy to postoperative radiotherapy for prostate cancer: a comparison of short-course versus no androgen deprivation therapy in the RADICALS-HD randomised controlled trial. The Lancet 2024;403:2405–15. 10.1016/S0140-6736(24)00548-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [6].Puri D, Pandit K, Choi N, Rose BS, McKay RR, Bagrodia A. Striving for Equity: Examining Health Disparities in Urologic Oncology. Cancers 2024;16:3559. 10.3390/cancers16213559. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [7].Esteva A, Feng J, Van Der Wal D, Huang S-C, Simko JP, DeVries S, et al. Prostate cancer therapy personalization via multi-modal deep learning on randomized phase III clinical trials. Npj Digit Med 2022;5:71. 10.1038/s41746-022-00613-w. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [8].Spratt DE, Tang S, Sun Y, Huang H-C, Chen E, Mohamad O, et al. Artificial Intelligence Predictive Model for Hormone Therapy Use in Prostate Cancer. NEJM Evidence 2023. 10.1056/EVIDoa2300023. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [9].Roach M, Zhang J, Mohamad O, Van Der Wal D, Simko JP, DeVries S, et al. Assessing Algorithmic Fairness With a Multimodal Artificial Intelligence Model in Men of African and Non-African Origin on NRG Oncology Prostate Cancer Phase III Trials. JCO Clin Cancer Inform 2025:e2400284. 10.1200/CCI-24-00284. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [10].Parker CTA, Mendes L, Liu VYT, Grist E, Joun S, Yamashita R, et al. External validation of a digital pathology-based multimodal artificial intelligence-derived prognostic model in patients with advanced prostate cancer starting long-term androgen deprivation therapy: a post-hoc ancillary biomarker study of four phase 3 randomised controlled trials of the STAMPEDE platform protocol. The Lancet Digital Health 2025:100885. 10.1016/j.landig.2025.100885. [DOI] [PubMed] [Google Scholar]
- [11].Shipley WU, Seiferheld W, Lukka HR, Major PP, Heney NM, Grignon DJ, et al. Radiation with or without Antiandrogen Therapy in Recurrent Prostate Cancer. N Engl J Med 2017;376:417–28. 10.1056/NEJMoa1607529. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [12].Pollack A, Karrison TG, Balogh AG, Gomella LG, Low DA, Bruner DW, et al. The addition of androgen deprivation therapy and pelvic lymph node treatment to prostate bed salvage radiotherapy (NRG Oncology/RTOG 0534 SPPORT): an international, multicentre, randomised phase 3 trial. The Lancet 2022;399:1886–901. 10.1016/S0140-6736(21)01790-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [13].Ilse M, Tomczak JM, Welling M. Attention-based Deep Multiple Instance Learning 2018. 10.48550/arXiv.1802.04712. [DOI] [Google Scholar]
- [14].Fine JP, Gray RJ. A Proportional Hazards Model for the Subdistribution of a Competing Risk. Journal of the American Statistical Association 1999;94:496–509. 10.1080/01621459.1999.10474144. [DOI] [Google Scholar]
- [15].Blanche P, Dartigues J, Jacqmin-Gadda H. Estimating and comparing time-dependent areas under receiver operating characteristic curves for censored event times with competing risks. Statistics in Medicine 2013;32:5381–97. 10.1002/sim.5958. [DOI] [PubMed] [Google Scholar]
- [16].Vickers AJ, Cronin AM, Elkin EB, Gonen M. Extensions to decision curve analysis, a novel method for evaluating diagnostic tests, prediction models and molecular markers. BMC Med Inform Decis Mak 2008;8:53. 10.1186/1472-6947-8-53. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [17].Gray RJ. A Class of K-Sample Tests for Comparing the Cumulative Incidence of a Competing Risk. The Annals of Statistics 1988;16:1141–54. [Google Scholar]
- [18].Tendulkar RD, Agrawal S, Gao T, Efstathiou JA, Pisansky TM, Michalski JM, et al. Contemporary Update of a Multi-Institutional Predictive Nomogram for Salvage Radiotherapy After Radical Prostatectomy. J Clin Oncol 2016;34:3648–54. 10.1200/JCO.2016.67.9647. [DOI] [PubMed] [Google Scholar]
- [19].Paszke A, Gross S, Massa F, Lerer A, Bradbury J, Chanan G, et al. PyTorch: An Imperative Style, High-Performance Deep Learning Library 2019. 10.48550/arXiv.1912.01703. [DOI] [Google Scholar]
- [20].Gerrard P, Zhang J, Yamashita R, Huang H-C, Nag S, Nhek S, et al. Analytical Validation of a Clinical Grade Prognostic and Classification Artificial Intelligence Laboratory Test for Men with Prostate Cancer. AI in Precision Oncology 2024;1:119–26. 10.1089/aipo.2024.0004. [DOI] [Google Scholar]
- [21].Bulten W, Pinckaers H, Van Boven H, Vink R, De Bel T, Van Ginneken B, et al. Automated deep-learning system for Gleason grading of prostate cancer using biopsies: a diagnostic study. The Lancet Oncology 2020;21:233–41. 10.1016/S1470-2045(19)30739-9. [DOI] [PubMed] [Google Scholar]
- [22].Huang W, Randhawa R, Jain P, Hubbard S, Eickhoff J, Kummar S, et al. A Novel Artificial Intelligence–Powered Method for Prediction of Early Recurrence of Prostate Cancer After Prostatectomy and Cancer Drivers. JCO Clinical Cancer Informatics 2022:e2100131. 10.1200/CCI.21.00131. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [23].Cha E, Lin Z, Lu J, Oliveira LD, Erak E, Mendes AA, et al. Histopathology-based Artificial Intelligence Algorithms for the Prediction of Prostate Cancer Metastasis After Radical Prostatectomy. European Urology 2025:S0302283825004956. 10.1016/j.eururo.2025.08.018. [DOI] [PubMed] [Google Scholar]
- [24].Ding C-KC, Shee K, Cowan JE, Ren Y, Tang S, Yang X, et al. External validation of a pathology-based multimodal artificial intelligence biomarker for predicting prostate cancer outcomes after prostatectomy. JCO 2025;43:5106–5106. 10.1200/JCO.2025.43.16_suppl.5106. [DOI] [Google Scholar]
- [25].Carrie C, Magné N, Burban-Provost P, Sargos P, Latorzeff I, Lagrange J-L, et al. Short-term androgen deprivation therapy combined with radiotherapy as salvage treatment after radical prostatectomy for prostate cancer (GETUG-AFU 16): a 112-month follow-up of a phase 3, randomised trial. The Lancet Oncology 2019;20:1740–9. 10.1016/S1470-2045(19)30486-3. [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.