The Problem with Natural Language Processing
The problem with natural-language-processing in medicine is that the medical record is often fictional. So the algorithm will be useless.
Image from Atul Gawande’s article “Why Doctors Hate Their Computers” in the New Yorker.
Developers and health institutions are spending countless resources to extract automated insights from medical records with Natural Language Processing software. They think they have forced doctors to automate their own jobs by writing everything down in an EHR. NLP can work when you have a real text; a court transcript, a book, or a voice conversation…
But most official medical notes are elaborate fabrications:
Copy pasted information: No one has time to rewrite clinical notes anymore so the majority of notes are copy-pasted from the day before. Behavior which can create what Dr. Abraham Verghese calls "the iPatient”.
Forced click-drag behavior: Clinicians are forced into click-drag behavior that does not fit a kinesthetic assessment and was never coded to fit medical diagnostics.
Written for malpractice lawyers: Clinical notes are written to stand up in court years later, and doctors practice “defensive medicine” to avoid the US malpractice industry. Clinical notes can be increasingly unrelated to decisionmaking under uncertainty.
A creative retelling for the best outcome, 'active' data: Real-time clinical decisions and assessments are being conducted on smartphones in text strings with photos, left out of the record completely and untraceable.
It is very dangerous to mistake a doctor's note for their work. Especially these days when the notes are so difficult to write. Developers are automating a fictionalized record, and the resultant algorithm will also be fictional.
We need black-box transcripts, 'passive' data first, a culture that stops blaming individuals and starts fixing systems second, THEN we can start creating algorithms that have some utility.
#ai #nlp #naturallanguageprocessing #healthcare #todreamalife #medicine #medicolegal #machinelearning #deeplearning
