As pretty much anyone who follows health data analytics knows, AI is proving to be a powerful tool for finding hidden or subtle trends in big patient data stores we already have. What may not be as obvious is that the growth in healthcare AI use is also changing how healthcare data is gathered and used in the first place.
For example, according to research by CB Insights, some healthcare organizations are adopting a “federated learning” approach to training AI tools, the report says. In this model, the majority of patient data never leaves hospital premises and never hits a central cloud server. Instead, the dataset is updated at the hospital using local data, with only updates from internal resources or other participating hospitals going to the cloud.
This approach is likely to kick off a new hospital infrastructure arms race, as using a federated learning model calls for every hospital to have the infrastructure and staff available to train machine learning tools. In other words, we’re talking about some potentially substantial costs here. On the other hand, this certainly won’t stop deeper-pocketed systems from making the investment, and that could lead to healthcare data management norms changing shape. Plus, the federated model can still be less expensive than trying to extract, transform, and load (ETL) model.
Another new technology trend driven by healthcare AI is the use of generative adversarial networks (GANs) to create training material for AIs tasked with analyzing images. According to the research, these GANs can create realistic fake images, which can include more-realistic artificial tumors in MRI images, which in turn gives machine learning algorithms a lot more to chew on.
In other words, if hospitals use this technology they can create billions of appropriately-faked tumor-displaying MRIs rather than the millions or hundreds of millions they might have now, which can support the development of much smarter imaging analytics AIs. According to NVIDIA research cited by the report, this approach will address two of the biggest problems in supporting machine learning for medical imaging, a small incidence of pathological findings and restrictions related to sharing patient data freely.
Now, as with any discussion of whiz-bang new technology, it’s important not to get too far ahead of ourselves, Even if they can afford to do so, I can’t see the majority of hospitals and health systems rebuilding to support machine learning technology just yet. And as interesting as approaches like generative adversarial networks might become, my sense is that they’re still at best at the research stage.
It’s also worth noting that we haven’t put some of the key intellectual infrastructure in place we’ll need to work together on larger health AI projects, including agreeing on best practices and identifying some standard use cases for different AI deployments. Yes, it looks like some imaging analysis approaches are maturing rapidly, but we’re just getting started with other AI-fueled health data analytics models.
Not only that, it seems likely that we’ll see other intriguing and probably transformational options emerge from AI experimentation. It’s probably far too soon for health IT leaders to put all their eggs in any healthcare AI basket.
Still, I don’t know about you, but I’m very excited to see iterative health Al models leapfrogging each other. While they may have been solutions in search of a problem previously, they are quickly becoming solutions, full stop, to some of the biggest problems we’ve got in healthcare. If any emerging health care technology seemed poised to fulfill its potential, healthcare AI does.