AI Tool Trained on Tumor Images Could Help Doctors Decide Who Needs Abiraterone for Prostate Cancer

If you or someone you love has ever sat in an oncologist's office weighing a treatment decision, you know the quiet anxiety of that calculation. Will this drug help enough to justify what it might cost your body? For men with very high-risk, localized prostate cancer, that question has real stakes. A new study suggests artificial intelligence may finally offer a clearer answer.

Standard treatment for high-risk, nonmetastatic prostate cancer combines about three years of hormone therapy with radiation. For patients whose cancer is classified as very high-risk, doctors often add abiraterone, a drug that blocks the body's production of testosterone-fueling hormones that cancer cells feed on. The STAMPEDE trials, a long-running international research platform, showed this combination meaningfully extends survival. But abiraterone is not a free lunch. Grade 3 or higher side effects occur in 37% of patients who take it, compared to 29% on hormone therapy alone, and cardiac, vascular, and liver problems force 13% to 29% of patients to stop treatment altogether.

That tradeoff has left doctors without a reliable way to know, ahead of time, which patients genuinely need the extra drug and which might be exposed to its risks for little benefit. Roughly 62,500 American men each year meet the clinical criteria for abiraterone intensification, based on Surveillance, Epidemiology, and End Results database estimates. Clinical measures like tumor stage, PSA levels, and Gleason score give some guidance, but they are blunt instruments. Two patients can look identical on paper and have very different disease trajectories.

How the AI Model Works

Researchers led by a team publishing in Annals of Oncology turned to a tool called multimodal artificial intelligence, or MMAI, which had already shown promise as a prognostic test for prostate cancer. Think of it as a very sophisticated pattern-reader. The model examines digitized images of tumor tissue under a microscope alongside standard clinical data, like PSA and tumor stage, and combines them to generate a single risk score. It is the medical equivalent of a skilled radiologist and a lab analyst working together, except the AI processes patterns across thousands of cases that no single human reviewer could hold in mind at once.

The team applied this tool retrospectively to 1,137 patients who had been randomly assigned, as part of two sequential STAMPEDE trials, to either standard hormone therapy or hormone therapy plus abiraterone. Using a threshold established in earlier work, the top 25% of MMAI scores sorted patients into a "very high-risk" group, while the remaining 75% fell into a "standard high-risk" category, despite all of them meeting the clinical definition of very high-risk disease going into the trial.

The results were striking. Among the 268 patients the AI flagged as truly very high-risk, adding abiraterone cut the risk of metastasis or death nearly in half, with a hazard ratio of 0.47. Five-year metastasis-free survival jumped from 62% with hormone therapy alone to 81% with abiraterone added. That is the kind of improvement that changes a patient's life expectancy and quality of life in meaningful ways.

For the other 869 patients, the ones the AI classified as standard high-risk despite meeting the trial's clinical very high-risk criteria, abiraterone's benefit nearly disappeared. Five-year survival rates were 82% versus 84%, a gap small enough that it may not justify the drug's toxicity burden for many of these men. The statistical test for this difference, called an interaction p-value, came in at 0.02, indicating the split between groups was unlikely to be due to chance. Notably, this pattern held steady whether patients had cancer that had spread to nearby lymph nodes or not.

It is worth pausing on what this means practically. Conventional risk criteria, the ones doctors use every day, lumped all 1,137 of these patients into the same "very high-risk" bucket. The AI tool found that three out of four of them were unlikely to gain much from the added drug. That is not a small refinement. It is a fundamentally different way of sorting patients, one that could spare a majority of men from unnecessary side effects while still protecting the smaller group who truly need aggressive treatment.

This study arrives alongside other recent findings that complicate the picture of treatment intensification in prostate cancer. The ENZARAD trial, which tested a similar strategy using a different drug called enzalutamide in a broader population of high-risk patients, found no significant survival benefit at all. Researchers noted that the ENZARAD population generally had lower-risk disease than the STAMPEDE cohort, with an 8-year risk of prostate cancer death around 3%, compared to roughly 19% in STAMPEDE. Several other large trials testing similar drugs have finished recruiting but have not yet reported results. Taken together, these studies underscore just how much risk and benefit can vary even among patients who share the same clinical label.

Why It Matters

Medicine has long struggled with this exact problem: broad categories that group people together based on convenient but imprecise criteria, while obscuring the real biological differences underneath. Precision oncology promises to close that gap, and this study offers a concrete, validated example of how it might work in practice.

The researchers describe their findings as a post hoc analysis of randomized trial data, meaning the AI model was applied after the fact to patients who had already been through the trials. That is an important caveat. It means this is strong evidence, not yet a finished clinical tool ready for every oncology clinic. Before this test becomes part of routine care, it will likely need prospective validation, where doctors use the AI score to guide treatment decisions in real time rather than analyzing it retrospectively.

Still, the implications are hard to overstate. For patients, a tool like this could mean avoiding months of unnecessary treatment and its side effects. For health systems, it could mean directing costly drugs toward the patients most likely to benefit. And for the broader field of oncology, it is a reminder that the next wave of progress may come less from new drugs and more from smarter ways of matching the drugs we already have to the patients who need them most.