Caffeinated Classification Criticism 2

August 15th, 2013 / By Ron Mills, PhD

The next time I ran into Dr. Y in the coffee shop, he started right up.

Dr. Y: We put a lot of time and effort into getting our EHR systems up and running. It was grueling. I sat on some of the work groups at the hospital. The EHR vendor needed us to tell them how we think, what we write down when. It was worse for the nurses, I suppose, but in the end we got it going. They tell me that most of what we put in there is represented as SNOMED codes. So why can’t you use those codes for billing instead of forcing us to hire coders and create a whole other ICD-10 system?

Me: The short answer is “maybe we can, someday.” The EHR vendors and the coding-and-reimbursement vendors, including 3M, are all working on it. But it will take decades, or at least years, as computers get faster and our NLP – Natural Language Processing – software gets smarter.

Dr. Y: No, I said why can’t you use the SNOMED codes for billing?

Me: Ah. A common confusion. I think it arises from the overloaded use of the word “code.” A “code” in my business is any string of letters and digits that stands for something, and doesn’t mean anything by itself, out of context. So instead of calling the representations in SNOMED and ICD-10 “codes” we should call them what they are.

Dr. Y: Which is?

Me: In SNOMED they are “concept identifiers.” In ICD-9 and ICD-10 they are  “diagnoses” and “procedures.” So your question becomes: Why can’t you use concept identifiers for billing instead of diagnoses and procedures?

Dr. Y: Wait. Some of those concept identifiers are diagnoses and procedures. All you would need is a table of correspondences between the SNOMED concept identifiers and the ICD-10 diagnoses and procedures, and you’re done.

Me: Not so fast. Such a table was built for ICD-9, and for all I know one has been built or is being built for ICD-10. That’s not the problem. The problem is that SNOMED is a nomenclature – a mechanism for talking about a patient – while ICD-10 is a classification – a mechanism for putting your findings into diagnosis cubbyholes and the work you do into procedure cubbyholes.

Dr. Y: But all the information – the findings, the procedures – is already in there …

Me: … but in context. “Possibly diagnosis this … not that … rule out the other one … give these pills to prevent something else … his father had Alzheimer’s …” The EHR is riddled with diagnoses. Somebody has to determine which ones to bill with. Same for procedures. Read an op report and you’ll see dozens of procedures for which, taken singly, there are ICD-10-PCS codes. But most of them are implicitly included in whichever procedure is agreed to be the reason the patient was in the operating room in the first place. Somebody has to put the discipline in.

Dr. Y: The discipline?

Me: Coding is a discipline and good coders are disciplined.

Dr. Y: !!! (looks at me like I’ve gone off the deep end)

Me: Let me try again. I’ll stick to hospital inpatient – I don’t do outpatient much and besides, outpatient billing is mostly based on CPT, which isn’t changing.

Dr. Y: (with one eyebrow raised) Okay.

Me: Broadly speaking, there are three steps. First, you do what you do and record it, on paper or in an EHR or whatever. Second, coders, with help from computers, apply their discipline to the record and produce a UHDDS abstract (Uniform Hospital Discharge Data Set) – a.k.a. ICD codes. They may have to pester you to amend or add to the record if what you have recorded seems to them to be inconsistent or incomplete, and in the end you sign off on their work. Third – this part is fully automated – computers take the UHDDS and compute a DRG and from that, payment.

Dr. Y: Right. So what I’m saying is eliminate the middle step. Put the “discipline” you are going on about into a computer program that computes the DRG from the EHR.

Me: Even if I could do that, you would still need the intermediate representation – the UHDDS, the ICD codes. The DRG definitions manual, the document that describes computing the DRG from the ICD coded record, is a few hundred pages long. A manual describing computing a DRG from an EHR – assuming far more rigorously standardized EHRs than we have now– would be tens, maybe hundreds, of thousands of pages long. Such a document is effectively part of the contract between the payer and provider. Lawyers have to be able to navigate it. Further, research into health care quality and efficiency depends on the disciplined ICD-level classification – it is hundreds of times more difficult doing analysis directly from an EHR. And given all that: ICD-10 makes a better disciplined representation of care than ICD-9 does.

Dr. Y: Fine. So ICD-10 becomes a necessary simplification of the journey from EHR to DRG that requires “coding” in the middle. I won’t have to be bothered with it. Sounds good to me. When will you have billing straight from the EHR finished?

Me: Ha ha. SNOMED and the other ontologies the EHRs use are languages – precise little languages, but languages nevertheless – so how soon we can have primarily computerized coding will depend on how fast our NLP expertise develops. I’d say, come back in twenty years and we won’t be having this conversation.

Dr. Y: (heading out the door) Okay, see you then.

Ron Mills is a Software Architect for the Clinical & Economic Research department of 3M Health Information Systems.