TMR podcast: Computers can’t read your clinical notes

1 minute read

Uncategorized

By

Staff Writers

Clinical notes are about as indecipherable to a computer as a doctor's handwriting is to the average non-pharmacist human

Human language is hard for AI to wrap its circuit boards around at the best of times.

But the typos, abbreviations, spelling mistakes, redundancy and messy structure of clinical notes makes it the worst-case scenario for text mining.

Lots of research teams are trying to solve this problem – but it’s proving much more difficult than anyone expected.

You can subscribe to The Medical Republic podcast on iTunes, Spotify or wherever you get your podcasts by searching for ‘The Medical Republic”.

Subscribe
Notify of
1 Comment
Oldest
Newest Most Voted
Inline Feedbacks
View all comments
Andrew Miller
Andrew Miller
4 years ago

This podcast illustrates one of the fundamental errors of NLP ‘researchers’. The medical environment is another culture with its own language, categories and mores which are inculcated over a decade of training. Firstly the audio – why do you overlay multiple tracks so that it is all unintelligible. For your information:

sarp pain ++ = very severe sharp pain
at rest 0000hours = started while at rest at midnight.
lupus = SLE

There are several problems. One problem is that medically untrained people think that they are able to develop algorithms for understanding medical language when they have no idea to start with what is actually written on the page. This activity is more akin to the decryption of hieroglyphics, not NLP. It’s NOT ‘natural’ language, it’s MEDICAL language.

It is a fundamental axiom in computer use that if humans can’t do something reproducibly and repetitively, computers are going to be more useless.

Another problem is that facts are not always facts. The history section in a pathology report is hearsay, not clinical reality. While the histopathology images are the primary data, the primary text is the description given by the pathologist, so that there are reliable and unreliable parts of all records.

Another problem is that “oils ain’t oils”, or rather “doctors ain’t doctors”. Those in the know will attest that there is little overlap in the language and concepts of psychiatrists and orthopaedic surgeons. Psychiatrists fundamentally map your problem into the DMS axes, while the orthopaedic surgeon is interested in the procedure. As an oncologist, I am fundamentally interested in the diagnosis and stage. This is why one EMR can’t possibly work for all doctors, and why one doctor can’t specify EMRs for doctors from other areas.

For what it’s worth, I doubt that the involvement of oncologists with IBM-Watson is in any way real involvement. My experience with and of data miners is that the doctor is necessary for success in grant application, but not then included in the rest of the project. I have a principal investigator status on a data mining project for more than 2 years (~0.1 FTE, I believe) but I have not had 1 minute in collaborative meeting with the data miners.

The other problem revolves around the use of so-so algorithms in commercial products, who pick up the algorithms early, spruik them as unbelievably good, take the money and then never be seen again.

Moreover I wonder whether doctors actually want data. I know they want to write their notes, but the contained data requires some extraction and doctors don’t want to do this.

0

End of content

No more pages to load

Log In Register ×