AI significantly improves risk classification for prostate cancer metastasis
Click Here to Manage Email Alerts
Key takeaways:
- Many men with prostate cancer received excessive treatment because of broad risk guidelines.
- AI can help oncologists personalize prostate cancer treatment.
An AI model classified risk for prostate cancer metastasis more accurately than standard guidelines, suggesting the approach could be used to reduce overtreatment or undertreatment.
More than half of patients classified as having intermediate risk for metastasis based on National Comprehensive Cancer Network guidelines actually had a low risk based on the AI model, retrospective study results showed. Nearly half of those characterized as high risk had an intermediate risk.
Jonathan D. Tward, MD, PhD, FASTRO, professor in the department of radiation oncology and leader of the Genitourinary Cancers Center at Huntsman Cancer Institute at The University of Utah., summarized his reaction to the results with three words: “Shocked. Shocked. Shocked.”
“I spent my career helping develop expensive tests using genomic sequencing and genomic profiles to try to get personalized risk signatures,” Tward told Healio. “If you would have told me that a computer can almost blindly look at an image of a few biopsy cores and come up with risk models that are better than all the promise of genomics and historic risk stratification the first try out of the gate, I would have said you were crazy. But this first foray into just digital histopathology already rivaled our best genomic signatures.”
‘Huge variability’ in risk groups
Anthony V. D’Amico, MD, PhD, professor of radiation oncology at Harvard Medical School and chief of genitourinary radiation oncology at Brigham and Women’s Hospital and Dana-Farber Cancer Institute, helped develop the D’Amico classification system for prostate cancer risk stratification in 1998.
The D’Amico system — which has been at the core of predicting metastasis risk since then — has been enhanced over the years. However, it is still based on three evaluations — PSA, Gleason score and how the prostate feels on a rectal exam.
Tward called the D’Amico system “revolutionary” but outdated.
“Those original risk groups were really designed around biochemical failure, which is, can we still detect PSA after treatment for prostate cancer?” Tward said. “The reason why that’s problematic is there are people who would meet this definition of biochemical failure who will never need a treatment again in their lifetime. It’s not a great oncologic endpoint.”
Tward shared an example of a 40-year-old man classified with unfavorable intermediate risk prostate cancer based on NCCN guidelines. Individuals in that stratification have a metastasis risk between 0% and 40%, Tward said.
Guidelines stipulate that, after that patient receives radiation, he should also get hormone therapy. Yet, many of those men likely would not need it, Tward said.
“It’s such a huge variability that it doesn’t make sense,” he said. “You don’t want to throw the book at someone with a 0% risk, and you don’t want to do too little for someone who’s got a 40% risk.”
In the late 1990s, clinicians did not have many prostate cancer treatment options. Men could receive surgery, radiation, or one or two hormone therapies.
Today, the FDA has approved 13 drugs that can be used to treat prostate cancer, as well as multiple radiation and surgical options, Tward said. The breadth of options should allow for more personalized care, he added.
“Current NCCN risk groups don’t do a great job of prognosticating more meaningful endpoints like, ‘Will you need additional treatment after your first treatment?’ ‘What is the risk for cancer spreading in spite of your first treatment?’ ‘What is your risk for dying?’” Tward added. “Those are much more meaningful endpoints. The current NCCN risk groups have tremendous heterogeneity on those more important endpoints, which is the impetus for these kinds of research solutions for personalizing care.”
Methods and results
Tward and colleagues used a multimodal AI to evaluate 2,486 men (79.5% white; median follow-up, 7.9 years; range, 6.1-17.1) with localized prostate cancer who had participated in one of eight randomized phase 3 trials.
Men received one or a combination of radiation therapy, androgen deprivation therapy and chemotherapy as part of their treatment.
Time to distant metastasis served as the primary endpoint. Time to death with distant metastasis served as the secondary endpoint.
Based on NCCN guidelines, the study population consisted of 30.4% men with low-risk disease, 25.5% with intermediate risk, and 44.1% with high risk.
The AI categorized the cohort as 43.5% low risk, 34.6% intermediate risk and 21.8% high risk. The AI would have moved 359 intermediate-risk and 63 high-risk men into the low-risk cohort.
Overall, AI recategorized 42% of patients.
The NCCN and AI groupings had similar 10-year metastasis risks (1.7% vs. 3.2%).
Patients at high NCCN risk had a 10-year metastasis risk of 16.6%.
The AI model broke that population down into low-, intermediate-, and high-risk groupings, determining 10-year metastasis risks of 3.4% in the low-risk group, 8.2% in the intermediate-risk group and 26.3% in the high-risk group.
“The NCCN guidelines say every man in the high-risk group, if they’re getting radiation therapy, should get hormone therapy with it for at least a year — usually a year and a half,” Tward said. “Men hate hormone therapy. The AI shows 46% of these [men] probably don’t have a risk for developing metastasis high enough to warrant giving them that long-term therapy ... and 5.9% have no more risk for developing metastasis than the people we currently put on active surveillance.”
Researchers acknowledged study limitations, including not evaluating the AI in patients with favorable vs. unfavorable intermediate risk.
‘The power of AI’
Tward said he does not know what the AI identified to reclassify patients, but he called that “the power of AI.”
“Imagine the AI is identifying 256 patterns when it’s interpreting [a prostate cancer] image,” he said. “Of those 256 patterns, maybe 25 are things we can get a human to recognize — maybe 10% — but 90% of the things it can see aren’t things a human can see. It’s a little scary. We don’t really know what it’s doing, we just know that it’s accurate. The reason we know that it’s accurate is because we know exactly what happened to these patients over time.”
AI can do much more in this space, Tward said, including determining whether certain drugs will work.
AI risk classification can improve, too. AI can identify patterns in data clinicians are currently “overlooking,” Tward said. Adding more information, such as MRIs and imaging, will improve accuracy, as well.
“The most amazing predictive algorithms still only get it right maybe 80% of the time, which is much better than an NCCN risk group, which will get you there 60% of the time,” Tward said. “The future is trying to get algorithms that are accurate 99% of the time.”
Collapse
Click Here to Manage Email Alerts