Yes, so in the San Antonio meeting we presented a paper looking at the development of an AI model to predict the risk of recurrence in patients, risk of late recurrence in hormone receptor positive breast cancer patients. And we developed and validated this model in the MSABP B-42 trial, which compared five years of letrozole versus five years of placebo in women that have been disease-free for five years after five years of endocrine therapy...
Yes, so in the San Antonio meeting we presented a paper looking at the development of an AI model to predict the risk of recurrence in patients, risk of late recurrence in hormone receptor positive breast cancer patients. And we developed and validated this model in the MSABP B-42 trial, which compared five years of letrozole versus five years of placebo in women that have been disease-free for five years after five years of endocrine therapy. So we developed actually three models. One was an image-only model based on scanning the H&E slide. The image and clinical model, what we call multi-model, is the image of the slide plus clinical pathologic factors. And then we developed a multi-model, multi-task model, which included the two previous factors, clinical and image, as well as the bone mineral density T-score, but as a target value. In other words, predicting not only risk of recurrence, but also bone mineral density T-score. And what we found when we compared the three models is that the multi-model, multi-task model has the better predictive discrimination. It resulted in a higher hazard ratio of the high versus low risk patients, and also resulted in a higher absolute benefit from extended letrozole therapy. And so what we found was when we applied that in B-42 to validate it, we saw that in the overall study cohort, patients with a high risk M3T, as we call it, multi-task model, had higher risk of recurrence compared to the low-risk patients. This was true for the placebo patients as well as the extended letrozole therapy patients. When we looked at it according to the MT3 model and nodal status, we found that the model categorizes about 29% of the node-negative patients as high-risk and about 18.7% of the patients as low-risk, of the node-positive patients as low risk. When we looked at the benefit in node-negative patients, whether you’re high or low, there wasn’t much benefit in terms of placebo versus letrozole. When we looked at the node-positive patients, the high-risk patients had significant benefit in the range of about 5%, 5.5%. But patients in the node-positive but low risk had very little benefit, about 1.5%. Although the hazard ratio is somewhat similar. So the relative reduction was the same. But because the low risk patients had such lower risk, the absolute benefit was less. So essentially, what we concluded from this, and then in addition to that, we validated this in an external data set from the TailorX study from 4,300 patients in TailorX that were disease-free after four and a half years of adjuvant endocrine therapy. We validated the model and found that essentially the multi-model, multi-task model discriminates in terms of outcome and categorizes patients in the high risk and low risk, with a significant difference between them, with a hazard ratio essentially of about 1.89. And the multi-model multi-task model was an independent predictor in TailorX, along with grade and tumor size. Some of the other predictors were not significant, such as Oncotype DX, for example, in TailorX. So essentially what we showed in this, developing and validating the model both internally in B-42 and externally in the TailorX, that this is a model that can predict low-risk versus high-risk patients, and as a result, may help us customize the treatment in terms of not treating potentially the low-risk and treating the high-risk, despite the fact that the relative reduction was not, so it wasn’t a predictive biomarker, but it was a prognostic biomarker that can help us selecting the most appropriate treatment, either extended therapy or not.
