We tested this idea in two clinical settings in breast cancer and also validated it in randomized trials. First, we looked at HER2 positive patients as defined by immunohistochemistry and FISH, and we applied an AI model trained to predict the HER2 status to these patients. And then we split the patients into those predicted as HER2 positive and those mistakenly predicted as HER2 negative by the model...
We tested this idea in two clinical settings in breast cancer and also validated it in randomized trials. First, we looked at HER2 positive patients as defined by immunohistochemistry and FISH, and we applied an AI model trained to predict the HER2 status to these patients. And then we split the patients into those predicted as HER2 positive and those mistakenly predicted as HER2 negative by the model. And we did this in the TAILORx randomized trial where patients, HER2 positive patients, were randomized to either receive trastuzumab, which is anti-HER2 therapy versus placebo. And we expect that these patients benefit from trastuzumab. But what we saw is that the benefit from trastuzumab concentrated in patients predicted as HER2 positive by the AI, whereas those predicted mistakenly as HER2 negatives did not benefit from trastuzumab even though they are HER2 positive patients. So this means that we can, using these model errors, we can identify patients where clinicians should have maybe closer monitoring or even intensify their treatment. In a second setting, we looked at breast cancer patients with low Oncotype scores, meaning below 26. And we applied our AI model that predicts the Oncotype from H&E to these patients and split them into those the AI predicted as Oncotype low and those predicted mistakenly as Oncotype high, meaning above 26. And we did this on the TAILORx randomized trial, where this group was randomized to either receive chemotherapy versus placebo plus endocrine therapy. And what we saw is that the patients that were predicted as high Oncotype risk by the model actually benefited from chemotherapy, even though their true Oncotype score was low. And importantly, this was also observed in postmenopausal women. This shows that discordances between morphology-based signals and molecular-based signals can reveal, can provide a functional state, functional signal of the tumor that can actually help clinicians in making better clinical decisions. What’s interesting is that we do not have to have the follow-up or outcome data or gene expression data of the patients in order to reveal this functional signal, which has huge practical advantages. I believe that these new observations and analysis can shape our view and interpretation in the future of AI predictions from histology images in computational pathology and oncology.
This transcript is AI-generated. While we strive for accuracy, please verify this copy with the video.
