There are a lot of elements involved in doing predictive analytics in healthcare effectively. In most cases I’ve seen, organizations working on predictive analytics do some but not all that’s needed to really make predictive analytics as effective as possible. This was highlighted to me when I recently talked with Frank Stearns, Executive Vice President from HBI Solutions at the Digital Health Conference in NYC.
Here’s a great overview of the HBI Solutions approach to patient risk scores:
This process will look familiar to most people in the predictive analytics space. You take all the patient data you can find, put it into a machine learning engine and output a patient risk score. One of the biggest trends happening with this process is the real-time nature of this process. Plus, I also love the way the patient risk score includes the attributes that influenced a patients risk score. Both of these are incredibly important when trying to make this data actionable.
However, the thing that stood out for me in HBI Solutions’ approach is the inclusion of natural language processing (NLP) in their analysis of the unstructured patient data. I’d seen NLP being used in EHR software before, but I think the implementation of NLP is even more powerful in doing predictive analytics.
In the EHR world, you have to be absolutely precise. If you’re not precise with the way you code a visit, you won’t get paid. If you’re not precise with how the diagnosis is entered into the EHR, that can have long term consequences. This has posed a real challenge for NLP since NLP is not 100% accurate. It’s gotten astoundingly good, but still has its shortcomings that require a human review when utilizing it in an EHR.
The same isn’t true when applying NLP to unstructured data when doing predictive analytics. Predictive analytics by its very nature incorporates some modicum of variation and error. It’s understood that predictive analytics could be wrong, but is an indication of risk. Certainly a failing in NLP’s recognition of certain data could throw off a predictive analytic. That’s unfortunate, but the predictive analytics aren’t relied on the same way documentation in an EHR is relied upon. So, it’s not nearly as big of a deal.
Plus, the value that’s received from applying NLP to pull out the nuggets of information that exists in the unstructured narrative sections of healthcare data is well worth that small amount of risk of the NLP being incorrect. As Frank Stearns from HBI solutions pointed out to me, the unstructured data is often where the really valuable data about a patients’ risk score exist.
I’d be interested in having HBI Solutions do a study of the whole list of findings that are often available in the unstructured data that weren’t available otherwise. However, it’s not hard to imagine a doctor documenting patient observations in the unstructured EHR narrative that they didn’t want to include as a formal diagnosis. Not the least of these are behavioral health observations that the doctor saw, observed, and documented but didn’t want to fully diagnose. NLP can pull these out of the narrative and include them in their patient risk score.
Given this perspective, it’s hard to imagine we’ll ever be able to get away from using NLP or related technology to pull out the valuable insights in the unstructured data. Plus, it’s easy to see how predictive analytics that don’t use NLP are going to be deficient when trying to use machine learning to analyze patients. What’s amazing is that HBI Solutions has been applying machine learning to healthcare for 5 years. That’s a long time, but also explains why they’ve implemented such advanced solutions like NLP in their predictive analytics solutions.
