Long-term androgen deprivation therapy with radiotherapy is a standard treatment approach for patients with high-risk localized prostate cancer. For patients with clinically very high-risk non-metastatic disease, results from the STAMPEDE platform have supported adding abiraterone-based intensification to long-term androgen deprivation therapy and radiotherapy in this setting.
However, not all patients may derive the same level of benefit from treatment intensification. Abiraterone can add toxicity, treatment burden, and cost, creating a need for biomarkers that can better identify patients most likely to benefit.
A new post-hoc biomarker analysis evaluated whether a multimodal artificial intelligence (MMAI) model using digital pathology and clinical variables could predict benefit from abiraterone in patients with non-metastatic very high-risk prostate cancer treated in two STAMPEDE phase 3 trials.
The original article, titled “Multimodal Artificial Intelligence Prediction of Abiraterone Efficacy in Two STAMPEDE Phase 3 Trials of Non-Metastatic Very High-Risk Prostate Cancer,” was published in Annals of Oncology on June 5, 2026.
Authors: C.T.A. Parker, H-C. Huang, E. Grist, L. Mendes, R. Yamashita, D. Croucher, A. Sachdeva, L. Murphy, V.Y.T. Liu, S. Santos Vidal, T. Todorovic, S. Lall, M. Goncalves, S. Thakali, A. Wingate, L. Zakka, D. Wetterskog, K. Nowakowska, C.L. Amos, STAMPEDE Collaborators, D.M. Berney, P.T. Tran, D.E. Spratt, M.R. Sydes, L.C. Brown, S.G. Zhao, P. Nguyen, N.W. Clarke, C.J. Sweeney, E.L. Stewart, M.K.B. Parmar, A. Esteva, N.D. James, and G. Attard.
Why This Study Matters
The STAMPEDE trials previously showed that adding abiraterone acetate plus prednisolone, with or without enzalutamide, to long-term androgen deprivation therapy-based standard care improved outcomes in high-risk non-metastatic prostate cancer.
Still, the clinical population labeled “very high-risk” is heterogeneous. Some patients have aggressive disease and may need treatment intensification, while others may have a lower risk of progression and could be exposed to additional toxicity with limited benefit.
The multimodal artificial intelligence model evaluated in this study was designed to integrate digitized hematoxylin and eosin-stained prostate biopsy images with routinely collected clinical variables, including age, prostate-specific antigen, and clinical tumor stage. The model had previously been validated as a prognostic biomarker. This analysis tested whether it could also predict which patients benefit most from abiraterone.
You can also read about Prostate Cancer Cure Rate on OncoDaily.
Study Design
This was a post-hoc analysis of randomized clinical trial data from two sequential STAMPEDE phase 3 abiraterone trials. The analysis included patients with non-metastatic clinically very high-risk prostate cancer who had all required data available for multimodal AI score generation.
Eligible patients had either local node-positive disease or node-negative disease with at least two high-risk features, including tumor stage T3–T4, Gleason score 8–10, PSA ≥40 ng/mL, or relapse with high-risk features.
Patients had been randomized to long-term androgen deprivation therapy or long-term androgen deprivation therapy with abiraterone-based treatment. The standard-of-care backbone included long-term androgen deprivation therapy and radiotherapy according to protocol recommendations.
The primary endpoint of this biomarker analysis was metastasis-free survival. Other outcomes included overall survival, prostate cancer-specific mortality, and time to distant metastasis. The locked multimodal AI model classified patients into two groups using a previously defined upper-quartile threshold:
- MMAI very high-risk: 268 patients
- MMAI standard high-risk: 869 patients
In total, 1,137 patients were included in the complete multimodal AI cohort, including 583 assigned to long-term androgen deprivation therapy and 554 assigned to long-term androgen deprivation therapy with abiraterone-based treatment. Median follow-up in the MMAI cohort was 6.1 years.
Main Results
In the overall MMAI cohort, adding abiraterone improved 5-year metastasis-free survival from 77% with long-term androgen deprivation therapy to 83% with abiraterone-based intensification. The magnitude of benefit differed by MMAI-defined risk group. Among patients classified as MMAI very high-risk, abiraterone was associated with a significant improvement in metastasis-free survival:
- HR: 0.47
- 95% CI: 0.31–0.70
The estimated 5-year metastasis-free survival increased from 62% with long-term androgen deprivation therapy to 81% with abiraterone-based treatment. In contrast, patients classified as MMAI standard high-risk had limited evidence of benefit from adding abiraterone:
- HR: 0.83
- 95% CI: 0.63–1.09
In this group, estimated 5-year metastasis-free survival was 82% with long-term androgen deprivation therapy and 84% with abiraterone-based treatment. The biomarker-by-treatment interaction for metastasis-free survival was statistically significant, with an interaction p-value of 0.02. A similar pattern was observed for overall survival and prostate cancer-specific mortality, but not for time to distant metastasis.
Node-Negative and Node-Positive Subgroups
The model’s predictive signal was observed in both node-negative and node-positive patients.
In node-negative patients, the MMAI very high-risk group had improved metastasis-free survival with abiraterone:
- HR: 0.45
- 95% CI: 0.24–0.84
The estimated 5-year metastasis-free survival increased from 67% with long-term androgen deprivation therapy to 87% with abiraterone-based treatment. In node-positive patients, the MMAI very high-risk group also appeared to derive greater benefit:
- HR: 0.48
- 95% CI: 0.28–0.82
The estimated 5-year metastasis-free survival increased from 57% to 74%. In contrast, MMAI standard high-risk patients showed no clear evidence of benefit in either nodal subgroup.
Role of Digital Pathology Features
Traditional clinical variables alone did not significantly identify patients with differential benefit from abiraterone. The study found no significant treatment interaction using clinical high-risk factors alone. An input perturbation analysis suggested that digital image features contributed most of the model’s ability to predict differential absolute treatment benefit from abiraterone, accounting for 77.4% of the model’s predictiveness. This supports the importance of pathology-derived information beyond standard clinical risk factors.
Safety Context
This biomarker analysis did not report new safety outcomes. However, the article notes that prior STAMPEDE data showed higher rates of grade 3 or higher adverse events with abiraterone-based treatment compared with long-term androgen deprivation therapy-based standard care. This remains clinically relevant because the purpose of the biomarker is not only to identify patients likely to benefit, but also to potentially avoid unnecessary treatment intensification in patients less likely to benefit.
Limitations
This was a post-hoc retrospective biomarker analysis of prospectively randomized clinical trial data. The sample size for this analysis was not predefined, although all patients with the required MMAI inputs were included. The study did not collect race or ethnicity data, limiting conclusions about generalizability across demographic groups.
The original STAMPEDE trials were conducted before widespread use of next-generation imaging such as PSMA-PET, meaning metastatic status was determined using conventional imaging. Although the predictive hypothesis and analytic framework were specified before outcome analysis, prospective validation will still be needed before the biomarker can be considered fully established for treatment selection.
Takeaway
This post-hoc analysis suggests that a locked multimodal AI digital pathology model may help identify patients with non-metastatic very high-risk prostate cancer who derive the greatest benefit from abiraterone-based treatment intensification.
Patients classified as MMAI very high-risk had a substantial improvement in metastasis-free survival with abiraterone, while patients classified as MMAI standard high-risk showed limited benefit. These findings support further prospective evaluation of multimodal AI-guided treatment personalization in high-risk localized prostate cancer, particularly to better balance treatment benefit against toxicity and treatment burden.
The full article was published in Annals of Oncology.