I’m a doctor: Here’s what I found when I asked ChatGPT to diagnose my patients

> So after my regular clinical shifts in the emergency department the other week, I anonymized my History of Present Illness notes for 35 to 40 patients — basically, my detailed medical narrative of each person’s medical history, and the symptoms that brought them to the emergency department — and fed them into ChatGPT.

It's quite shocking that the doctor would openly admit to violating HIPAA in such a brazen way.

HIPAA is incredibly broad in its definition of protected health information: if it's possible to identify an individual from data even through statistical methods involving other data that a third party might already conceivably possess, it's considered protected. It's inconceivable that the doctor would be able to sufficiently anonymize the data in this capacity and still provide enough detail for individual diagnoses.

There are processes for anonymizing data to disclose for research purposes, but they're pretty time-intensive, and no ED would allow a doctor to do it by himself, nor would they provide that turnaround in just "a couple of weeks". And the end results are a lot less detailed than what's needed for individual diagnoses like these.

I really wonder what the hospital will say if and when they see this post. Given the timeframe and details described in the post, it's really hard to believe that they signed off on this, and hospitals don't take lightly to employees taking protected and confidential data outside their systems without proper approval.

EDIT: It looks like this doctor works at a for-profit, standalone acute care clinic, rather than a traditional ED at a hospital, so my statement that hospitals don't take lightly to this stuff doesn't apply. The law still applies to for-profit standalone emergency care, but they tend to play fast and loose with these things much more than traditional health networks.

But it also gets around the common misdiagnoses for chronic conditions. It has a great description of Long Covid and ME/CFS for example whereas your typical Primary care is going to dismiss that patient with a Psychology diagnosis as is happening daily across the entire western world. Its less biased but its not going to find the rare things especially where the patient has missed something important.

Its a mixed bag just like it is with software. If you ask it to solve something simple it often does a decent job, but something complex and its confidently wrong. It doesn't show the self doubt of expertise that it needs to be a reliable tool yet it still requires the user has that expertise to be able to save time using it.

Some of the best performances I've seen out of ChatGPT are essentially "junior programmer" level. But it still requires clear instructions and close supervision. But GPT's training data includes GitHub, and it's used to power Copilot. It has arguably been trained to be a programmer.

In less familiar domains, like law or medicine, GPT has presumably undergone very limited training and tuning. It's essentially an "internet lawyer" or an "internet doctor."

In domains like this, it simply can't provide zero-shot professional results. Not with the current training data sets, and not with the current model performance.

Of course, we have no idea how quickly this gap will be closed. It might be 6 months or it might be 6 years. The future is looking deeply weird, and I don't think anyone has even begun to think through all the implications and consequences.

He probably didn't have any expectations. Merely experimentation and observation and maybe a bit of word of warning after seeing the results for anyone thinking of using it in ways it shouldn't be.

Back in 2004, took a seminar in college regarding Decision Support Systems and how they a manager or doctor could ask it a question and get a response to help them make a decision. Went to the doctor couple a years ago and he charged $300 to google search the symptoms. No thanks.