How can I use ChatGPT as a Doctor?

Initial Warnings and Advice

ChatGPT is currently not HIPAA compliant (though if you are a super motivated health tech startup, you can build an enterprise service on top of Microsoft Azure’s API and that would be HIPAA compliant). So, pasting your patients’ medical records into ChatGPT is a big no-no under the HIPAA Privacy Policy. While ChatGPT is not HIPAA compliant, you can be proactive with your own personal data and opt out of sharing it with OpenAI by completing this form. So at the very least, your data is not being used to improve their model’s performance and you are limiting your exposure.

**Please note that manually redacting your patient’s medical information is generally insufficient for HIPAA, as the protected health information qualification of “any other unique characteristic of a patient” has a larger scope than most realize. This is because computer algorithms are very sophisticated at re-identifying patients from data that we would think is anonymous.**

‍

Ok, so now that we have adequately warned you not to use your real patients’ data (seriously), here are some tips for writing an effective prompt to make the most of ChatGPT! In general, you want to be concise and have a clear goal with your requests; the more direct you are, the more likely you will get a good response. And use as much context as possible to help the model understand what you are looking for.

‍

Example Clinical Ideas for Prompts

In the past few years, patient messages in electronic health records (EHRs) have increased dramatically. What if we alleviate this problem through ChatGPT by responding to patient communications. As you can see below, ChatGPT is able to provide a response, and then you could edit the message quickly as needed.

Let’s say you are a primary care physician and your patient arrives for your visit for incontinence after her recent pregnancy, and you want to refer them to a urologist. Let’s use ChatGPT to automate a referral form for your patient with their past medical history and reason for referral. Obviously you might need to format the content to a specific layout, but this could make your process faster.

Let’s say that a patient arrives with pneumonia with a history of diabetes and hypertension. You want to use ChatGPT to convert medical language to ICD-10 codes and provide guidance on how to narrow down within the codeset. Let’s provide ChatGPT a quick snapshot of our patient’s medical history and see the recommended codes provided.

Perhaps you are a resident physician that is newer in their medical career, and you are looking for more research on a particular disease, especially when this is a new rotation in a speciality like neurology. Or maybe you just want to know what is the medical research around a disease because medicine is an ever evolving field. You could use ChatGPT to ask relevant medical questions and get a faster response than Google might provide.

And finally, and the use case we at Abstractive Health are the most interested in, summarizing medical notes. ChatGPT is able to provide a succinct summarization of long form text. Again, entering real patient charts into ChatGPT is not recommended as it would violate HIPAA. You should check out our main website if you are interested in this example!

‍

Some Challenges with using ChatGPT as a Doctor

You will have to copy and paste text from your EHR system into ChatGPT. Don’t worry, this problem will most likely be solved soon! And even though you redacted your clinical information, it’s always challenging to know if it truly HIPAA compliant. It’s probably safest to use a company that guarantees an NLP solution as being 100% HIPAA compliant! Any directly integrated EHR solution would be HIPAA compliant.

‍

The service could go down right when you need it. You might need a paid subscription with ChatGPT plus. Currently the free version of the service is not reliable, especially at peak hours as the service is throttled. So if you become too reliant on the free version, you might be painfully reminded of your monotonous clinical workflows and the need to complete them manually when you thought you automated away that annoyance!

‍

ChatGPT is way too chatty. You are already too busy reading content as a doctor, and now you have to add on top of that an AI model that just doesn’t know how to be succinct. ChatGPT just has a personality trait of talking just a tad too much. But you can fix this with some good prompt design. Check out the example here:

It’s not specialized in clinical knowledge, so it might get some information wrong. Factuality is a concern with large language models and a lot of research is going into the space. Companies and researchers are doing some great work at making strides to improve the models. You should make sure to manually review any response and correct as needed; you are responsible for anything ChatGPT writes for you. That said, ChatGPT was shown to be able to pass the US Medical Licensing Exam, so the accuracy is really good even in healthcare.

‍

ChatGPT might decide that your request is too risky, so it will avoid answering your question all together. Most of the time, you just need to alter your prompt a tad so ChatGPT understands you are a physician.

Abstractive Health

Our startup is looking to incorporate some of these functions of ChatGPT, starting with summarization of medical charts, into a specialized healthcare version that would integrate directly into the EHR. We will use a different language model for the summarization and it would be HIPAA compliant, and a chatbot would know information about your patient automatically through the EHR integration. If you are interested in how generative AI is changing healthcare, you should subscribe to our newsletter!

‍