Androgen deprivation therapy (ADT) is a form of hormone therapy that, when added to radiotherapy, consistently improves cancer control in men with localized prostate cancer. Multiple clinical trials conducted since the 1980s have confirmed this benefit.
However, ADT carries significant side effects including hot flashes, loss of libido, erectile dysfunction, muscle loss, increased body fat, osteoporosis, and potential harm to heart and brain health. For many patients, these harms meaningfully reduce quality of life.
The critical problem is that the majority of men with localized prostate cancer never develop distant metastasis even with radiotherapy alone -- meaning they would not benefit from ADT but would still experience all its side effects. Despite this, there are no validated tools to identify which specific men actually benefit.
Current treatment guidelines recommend ADT based on prognostic NCCN risk groups and Gleason grading, but these tools were designed to predict outcomes, not to predict who will specifically benefit from ADT. This represents a major unmet need for personalized treatment decision-making.
Researchers developed a multimodal AI (MMAI) predictive model using data from five NRG Oncology phase III randomized clinical trials. In total, 5,727 patients with pre-treatment prostate tissue slides were included, providing one of the largest training datasets ever used for a prostate cancer AI tool.
The model was designed specifically as a predictive biomarker -- meaning it would identify which patients derive differential benefit from ADT -- rather than simply predicting who has worse prognosis. This distinction is clinically critical, as some patients may have poor prognosis but still not benefit from ADT.
Four trials (NRG/RTOG 9202, 9413, 9910, and 0126) were used for training, while the fifth trial (NRG/RTOG 9408) -- the largest published trial comparing radiotherapy with or without 4 months of ADT -- was reserved as a completely independent validation set with 14.9 years of median follow-up.
The primary endpoint used to train the model was time to distant metastasis rather than biochemical recurrence or overall survival. This choice was deliberate: ADT suppresses PSA regardless of tumor biology (making biochemical recurrence unreliable), while overall survival in this population is mostly driven by non-cancer causes of death.
The model was built in two components. The first was an image feature extraction model trained using self-supervised learning (SSL) on over 2.5 million 256x256-pixel tissue patches from digitized hematoxylin and eosin (H&E) stained biopsy slides. This step learned to recognize tissue patterns without using clinical outcomes data.
The second component was a downstream multimodal predictive model that combined the extracted image features with clinical variables. The architecture was trained to optimize the difference in ADT benefit between patient subgroups, outputting a continuous score called 'delta' that quantifies predicted treatment benefit.
Patients were classified as predictive model positive (predicted to benefit from ADT) or predictive model negative (predicted to derive minimal benefit) using the 67th percentile of delta scores as the threshold cutoff. This cutoff was chosen to maximize the separation between subgroups while keeping groups large enough for clinical use.
Histopathology image features -- including Gleason scoring patterns and additional imaging features learned by deep learning -- contributed over 86% of the model's prediction weight, confirming that the tissue slide itself contains rich biological information beyond what pathologists currently extract by eye.
In the validation cohort of 1,594 patients, the AI model classified 543 patients (34%) as predictive model positive -- the group predicted to benefit significantly from ADT -- and 1,051 patients (66%) as predictive model negative. Despite this different ADT benefit prediction, the two groups had similar baseline PSA levels, tumor stage, and NCCN risk group distribution.
Among predictive model positive patients, adding ADT to radiotherapy dramatically reduced the 15-year risk of distant metastasis from 14.4% to 4.0% -- an absolute reduction of 10.5 percentage points (sHR = 0.34, p less than 0.001). This is a very large treatment benefit.
Among predictive model negative patients, ADT provided essentially no measurable benefit: the 15-year distant metastasis rates were virtually identical between the ADT and no-ADT groups (6.9% vs 7.4%, sHR = 0.92, p = 0.71). The interaction between the predictive model and treatment was statistically significant (p-interaction = 0.01).
The predictive model positive group gained 0.8 years of distant metastasis-free time over 15 years by adding ADT, while the predictive model negative group gained only 0.1 years -- a clinically meaningful difference that confirms the model's ability to identify who truly benefits.
The benefit of the predictive model extended beyond just preventing distant spread -- it also identified patients who avoided dying of prostate cancer. Among predictive model positive patients, ADT reduced the 15-year prostate cancer-specific mortality from 12.7% to 2.6% (sHR = 0.28).
In predictive model negative patients, prostate cancer-specific mortality was nearly identical whether or not patients received ADT (5.3% vs 6.5%, sHR = 0.74). This means that for most intermediate-risk patients, adding ADT does not save lives from prostate cancer.
The absolute difference in 15-year prostate cancer mortality was 10.2 percentage points in predictive model positive patients versus only 1.2 percentage points in negative patients -- a striking contrast that reinforces the clinical value of identifying the right subgroup.
Notably, these cancer mortality results were achieved even though the model was not specifically trained to predict cancer-specific death -- it was trained only on distant metastasis. This suggests the model is capturing genuine tumor biology that drives both metastasis and cancer-related death.
The most impactful clinical finding is that over 60% of intermediate-risk patients enrolled in the validation trial could potentially be spared ADT based on the AI model -- meaning they would avoid significant side effects without compromising their cancer outcomes.
Current clinical guidelines recommend ADT for intermediate-risk patients, yet this study shows that NCCN risk groups are not predictive of ADT benefit. A patient classified as intermediate-risk is equally likely to be a predictive model positive or negative case, meaning risk group alone is not sufficient to guide this treatment decision.
The study also highlights an important philosophical shift in oncology: prognostic tools identify patients at higher risk, while predictive tools identify patients who will respond to a specific treatment. These are distinct questions, and this AI model was designed specifically to answer the predictive question.
The model is currently commercially available through the Artera laboratory as a test that physicians can order for prostate cancer patients. This represents a path toward implementing AI-driven, individualized ADT decision-making in routine clinical practice.
A key strength is that the model was validated in a truly independent, prospective randomized trial that was not used in training. This is the gold standard for validating predictive biomarkers and guards against overfitting -- a common pitfall where AI models perform well in development but fail in new patients.
The validation cohort included 20% African American patients, higher than the national average for prostate cancer diagnoses. This diversity strengthens confidence that the model's predictions apply broadly across patient populations, though the study was underpowered to assess predictive performance specifically within this group.
Important limitations include the lack of some modern clinicopathologic data (such as percentage of Gleason pattern 4 or positive biopsy core information) and the fact that this was not a prospectively designed model trial. Advanced molecular imaging was not used in the era these trials were conducted.
The study is also an important proof of concept that deep learning models trained on routine H&E tissue slides -- the same slides used in standard pathology -- can capture treatment-predictive biological information that goes far beyond what is captured by current Gleason grading or clinical risk stratification alone.
This study demonstrates for the first time that an AI model derived from digital pathology images and clinical data can identify which prostate cancer patients will and will not benefit from short-term ADT -- a treatment decision that has been guided by broad population-level statistics rather than individual biology.
The findings support a new treatment paradigm: rather than recommending ADT to all intermediate-risk patients, clinicians could use this AI model to reserve ADT for the approximately one-third of patients who will receive large, clinically meaningful protection from distant spread and cancer death.
This approach could reduce overtreatment and the associated physical and psychological burdens of ADT for the majority of intermediate-risk prostate cancer patients, while ensuring that those who genuinely benefit from hormone therapy receive it.
More broadly, this research illustrates the transformative potential of AI applied to routine clinical data -- not to replace pathologists or oncologists, but to extract richer biological information from existing tissue samples and guide individualized treatment decisions in ways that were previously impossible.