Big data being filtered into an illustrated brain

Common Miscon-ceptions in Healthcare AI

Misconceptions abound in the developing AI market that have been hard to define, thanks in part to the surrounding hype. Of the 57 organizations KLAS interviewed about using AI in their organizations, even those who saw success with AI technology experienced some roadblocks. In those organizations’ experiences, we find some common points that are helpful for those organizations on the cusp of purchasing or implementing an AI tool.

For me, these misconceptions put into perspective the underlying work both providers and vendors need to put forth in order to see outcomes from AI. Here are two of the most common misconceptions from the Healthcare AI 2019 report.

Misconception: Building models is the most time-consuming AI task

Providers and vendors often think that the most time-consuming piece of an AI project will be building out the model. But, on the other hand, they often underestimate the time it takes to prepare the data. The fix? Set proper expectations for how much effort it takes the customer to get data ready to test. Healthcare data is, understandably, pretty complex. Providers have a lot of unstructured data from clinical notes and other documents, and sometimes multiple EMRs are involved. Sometimes, getting accurate, comprehensive analysis also depends on the breadth of data including different systems that host financial, operational, and other patient related data. Where you pull data from can sometimes be pretty archaic. But, if key factors or value are missing, the model can break down.

All of this translates to a process that can be very manual. I think customers are aware of the importance of these steps, but they still do not realize how much time and energy it takes. So what does this mean? Firstly, it means that helpful AI vendors will have good data preparation services wrapped into their AI offering. Secondly, good customers will take ownership to understand the data and have dedicated resources to make sure the right data is ready for the model to test and utilize.

A Senior Data Architect explained that, “If you don’t have clean data to push through the AI tool, you aren’t going to get good results. If you cannot hit that threshold, then you might need to invest more on the data governance side. It is on us to make sure that we are not giving the tool sparse data and that we are bringing in data points that are consistent. The tool isn’t magic; we still have to take responsibility to cleanse our data, but it’s pretty close. If you give your vendor good enough data, then they can shape it in a way that is useful and appropriate for each of the models that they put together.”

Misconception: Our organization will jump at the chance to leverage AI tools

Organizations frequently feel that staff would love the chance to use AI in their work. While it is true that AI tools can potentially offer great insights that may enable providers to course correct or do something different, it is tempting to believe that the organization can just change instantaneously. Culture shifts really do take time.

It can be challenging for provider staff, including clinicians, to trust recommendations generated by AI and take meaningful actions. Clinicians are coming from a place of clinical experience. They may consider an interesting recommendation from the tool, but they may also question it and instead rely on their own experience and training. Culturally, that resistance needs to be overcome with the explainability of AI, training, and early staff buy-in for the tool.

Sharing success stories can also be a useful way to introduce an AI tool to staff. A Vice President of Clinical Systems told us, “One of the biggest challenges for artificial intelligence in an organization is that cultural shift to really trust that a computer is giving information and helping draw insight that I may otherwise not be able to do without it. But now they frame it in a way where it makes sense to people. Also, there are more use cases and more information and publications about the vendor we went with, so people are just generally more comfortable with it today than they were years ago.”

To learn about some other common misconceptions and ideas for overcoming them, I recommend reading the full report. In my next and last post, I will discuss from our research some of the best practices that both vendors and providers implemented to ensure success.




     Photo cred: Adobe Stock, Lee