[telchar]

I do NLP with medical notes. What I have observed is that the redundancy between the note and information available elsewhere in the EHR in structured formats is pretty high, as is the amount of boilerplate (e.g. section headers and list templates that may not be filled in).

This makes machine learning using the notes difficult since the content is so muddied up. It's far from impossible to do useful things with them but there is a lot of noise. Still, some things require us to look at the notes. For these things we would much rather have them than not.

While it's preferred to have the information entered in a structured way, doctors find that more of a hassle than entering in the data free-text and it's also no good for retrospective analysis where we didn't know several years ago we would be interested in something and so no structured field existed.

[abrichr]

Interesting insight! Where do you work?

[telchar]

I work for a hospital/health system.