[ialyos]

The article is simply wrong.

I know this because I worked as an ML engineer at an extremely successful company that automated medical coding using deep learning.

The confusion stems from conflating a "perfect solution" with a "human augmented" one.

90% of coding cases are trivial, have low value and can be done by a model. 10% are really subtle and need human expertise.

That's fine. You can make a billion dollar company on low hanging fruit. I think it's best not to conflate the perfect solution with a very good solution.

[javadocmd]

You've not refuted the article so much as pointed out a corner case the author didn't address in which ML is a good fit. Your example, using ML to perform the medical coding function, is using a data source (in this case the EMR) for one of the purposes for which it was explicitly designed and for which it is (arguably) non-deficient. That is a realm not doomed to failure.

The realm doomed to failure is using a data source for a completely oblique purpose for which it is horribly distorted. Namely, the purpose of optimizing individual and public health by discovering guidelines and treatments, diagnosing illness, and delivering optimal care.

(Of course medical billing as an enterprise shouldn't even exist, but that is another topic.)

[ialyos]

Thanks for the nuance. I completely agree with how you've framed the situation.

[sjg007]

Medical coding is just billing right? You match doctor notes to ICD-10 codes. Seems reasonable.

[nradov]

Medical coding is mainly billing with ICD-10 codes for diagnoses and CPT + HCPCS codes for procedures. However, there is also non-billing clinical coding for things like LOINC, SNOMED CT, and RxNorm.

[not2b]

What was was the criterion that you were optimizing? Currently hospitals try to assign codes in a way that maximizes payouts from insurance companies while avoiding straight up lying in a way that could cause them problems. So they'll handle that 10% by choosing the codes with the bigger payout.