Code explosion: Can current NLP technology handle ICD-10?
By Mark Morsch and Brian Potter, July 3, 2012
How making the jump from ICD-9 to ICD-10 can be less intimidating to coding professionals.
Whenever the regulation goes into effect — currently proposed for October 1, 2014 — the expectation is that ICD-10 will be overwhelmingly complex for coding professionals. The sheer increase in the number of new codes in the system — a five-time increase in diagnosis codes and a 19-time increase in procedure codes — may be daunting to even the best of coders. However, most do not realize that the code explosion comes from a handful of additional attributes that are being applied to codes they already know. Hospitals and health networks increasingly rely on computer-assisted coding (CAC) systems driven by natural language processing (NLP) technology, but not all NLP engines are created equal.
Here are two things providers need to look for in NLP to ensure it can manage the demands of ICD-10:
Compositionality — Suppose someone told you that you had to learn 10,000 actions, and then all of a sudden they increased that to 30,000 actions. That sounds bad. But, if they then tell you that all you really need to know is whether those same 10,000 actions happened left, right or unspecified, then it seems less overwhelming. This is called factoring out laterality, and it is an aspect of compositional approaches to NLP. Compositional approaches identify complex coding concepts by recognizing and combining their simpler pieces of meaning. This approach already exists in ICD-9 solutions and will simply be applied more frequently across the ICD-10 code base. If providers have a CAC solution with an NLP engine that can already recognize laterality for ICD-9, this is readily translatable to ICD-10. More than 24,000 ICD-10 code distinctions can be taken off the table just by factoring out laterality — reducing the effective code count dramatically in one fell swoop. You can let that sigh of relief go now. Laterality is just one of many concepts that can be factored out of the code bases in a compositional approach.
Robust medical concept knowledge base — Coders now recognize left and right, the concept of a lung, the concept of a joint and the concept of a removal procedure, but medical documentation doesn't always use the same words to describe these things. For example, with left and right, someone might abbreviate LT and RT in their documentation, lung may be pluralized as lungs, knee might be specified for a joint condition and removal might be inferred from a specific procedure, such as cholecystectomy. Saying coders recognize any of the base concepts, such as laterality, doesn't mean they (or their NLP engine) can get away with just knowing a few simple words. Your NLP has to be driven by a robust set of dictionary resources that provide information about medical terms and how they relate to each other. This is particularly important with body-part terminology. A huge percentage of the increase in procedure codes comes from greater specificity as to exactly where a procedure is performed.
There is an assumption that since an NLP engine automates what a coder is doing (at least on the surface), that if a coder is going to have a more difficult time with the number of ICD-10 codes that exist, so will the NLP engine. This is not necessarily the case, and knowing what your NLP engine is capable of now in ICD-9 will say a lot about how it's going to be able to manage ICD-10.
About the authors
Mark Morsch is vice president of technology and Brian Potter is manager of NLP innovation at OptumInsight. For more information, go to www.optuminsight.com.
Tags: :: Online Only Features ::