'AI Model Stratifies Late Recurrence Risk in HR-Positive Breast Cancer'

 At the San Antonio Breast Cancer Symposium, researchers presented findings on Clarity BCR, a multimodal multitask deep-learning algorithm that aims to estimate late distant recurrence risk in hormone receptor (HR)-positive breast cancer.

In this MedPage Today video, Eleftherios P. Mamounas, MD, MPH, of the Orlando Health Cancer Institute in Florida, discusses how deep-learning models developed in the NSABP B-42 trial -- and externally validated with the TAILORx trial data -- help identify patients at higher risk for late distant recurrence and those most likely to benefit from extended endocrine therapy.

Following is a transcript of his remarks:

In the study we presented in San Antonio, we looked at ways to predict better late distant recurrence for patients with estrogen receptor [ER]-positive breast cancer, which as we know have a significant risk for late recurrence.

So we looked at one of the studies, a clinical trial in NSABP B-42, that we evaluated 5 years of adjuvant letrozole therapy versus placebo for women that had ER-positive breast cancer, and they were disease free after 5 years of endocrine therapy with tamoxifen/aromatase inhibitor.

So in that study where we saw a small but statistically significant benefit in favor of letrozole, and because of this modest benefit, we wanted to be able to identify better whose patients have higher risk or low risk for late distant recurrence. So which patients potentially may benefit more or less from the extended endocrine therapy.

So using artificial intelligence, we develop three deep-learning models, one based on H&E imaging results, only scanning H&E slides. The second, what's called multimodal model, which included the H&E information as well as clinical pathologic factors. And the third model that we call multimodal and multitask, that has an outcome also included bone mineral density [data]. So included all the other information plus the prediction of more [bone] mineral density.

So first we developed and validated internally these models in the B-42 trial. And in that trial it shows us that the model, particularly the multimodal multitask models, we call it M3T, was the best model in predicting distant recurrence, late distant recurrence. So in B-42, that model essentially discriminated patients at a high risk or low risk, based on a 50/50 split. And this was true for patients that received either letrozole therapy or placebo.

And so these findings also then looked at prediction of benefit from extended letrozole therapy. And what we found was that the higher-risk patients appear to have higher absolute benefit from extended endocrine therapy in the range of 5-6% versus the low-risk patients had only about 1-1.5% benefit. And even if the hazard ratios were very similar -- so relative reduction was similar because of the lower risk in the low-risk group -- the absolute benefits were much smaller.

So we then externally validated that in the TAILORx study. And this, as we know, is the largest study that included node-negative patients [who were] HR-positive. The B-42 included node-negative and node-positive patients, but the TAILORx was mostly all node-negative patients. So we applied the same multimodal multitask, the M3T model in TAILORx, and saw a significant discrimination. Again, separated patients at high risk and low risk with a hazard ratio, in this case about 1.8%. And then in an independent multivariate analysis, the M3T model was an independent predictor along with grade and tumor size.

So essentially what we demonstrated in the study was that we can predict risk of late recurrence with an AI [artificial intelligence]-based model, the M3T model, and we externally validated that in the TAILORx. There wasn't much predictive information in terms of extended letrozole therapy benefit. But by finding patients that are low risk, they received less absolute benefit versus high-risk patients who received more absolute benefit.

https://www.medpagetoday.com/meetingcoverage/sabcsfuturefocus/119037