Using artificial intelligence to link genomic data to clinical outcomes in cancer care across institutions without compromising protected health information is challenging.
To both improve cancer care delivery and ask research questions. We need data on clinical outcomes. And one long-standing challenge has been that um these outcomes data are not recorded anywhere in the electronic health record except in the free text of imaging reports and oncologist notes, we and others have tried to tackle that problem by developing A I models or natural language processing models that can uh essentially automate that task of uh getting clinical outcomes data. Like is the cancer improving or responding? Where is it metastatic to out of the electronic health record? Um One of the challenges with making that Generali Zable is the fact that um anytime one trains an A I model on a data set, there's a risk that that A I model might memorize the data on which it was trained. And if we're talking about private protected health information that then can prevent us from being able to share these trained A I models. And so I think the the key take home is the the fact that this method may enable us to deploy A I at a broader scale without taking risk with the protected health information that might have been used to train the A I models uh in order to create large data sets for uh fuel and cancer research and also uh enable cancer centers and health systems uh to ask questions like who at our cancer center had progressive cancer yesterday. In order to inform cancer care delivery interventions.
