Prompt Engineering in Healthcare: Why Your Hospital Needs an Expert

Variability Is a Feature, and a Risk

An underrated problem in prompt engineering is using AI too much. It's important to define your scope, especially in healthcare.

What's important to understand is that AI is fundamentally random. This is a feature, not a bug. After lots of experimentation, LLM engineers found that if a Large Language model always chose the single most likely next word, the result would be dull or incoherent. It's like repeatedly pressing the middle suggestion on your phone's keyboard, it will create "sentences” but they won't mean anything.

As a prompt engineer, knowing this means you should never use AI to generate text that always has to be the same.

For example, if you want to have a disclaimer at the bottom of your letter, that has been approved by your legal team: Adding this into your assistant is absolutely the wrong way to go about this. Eventually, your assistant will eventually change the phrasing you asked for it to use. And while normally, we'd expect work to be signed off and verified by a clinician, the exact phrasing of a legal disclaimer should be is outside their realm of competency and checking would create extra work for the clinician (which defeats the whole point!).

This is similar to a related issue: formatting.

Any kind of formatting you need to go into a letter or a document should have as little to do with AI as possible. Theoretically, you could have your assistant export a completely formatted PDF for you to send out straight away. But it will always be inconsistent. A full PDF file has a lot of data, that can also change on a whim, that’s why you need to know exactly when your Assistant’s text begins and ends. Will you have headings? A sign off? A footer? All of these need to be known before you begin. Letters especially have a lot of different kinds of formats that could be applied to them, and assistants like to add things like footers, today's date, hallucinated addresses, etc.

Beginning this process of understanding your scope sooner, will help you get this work done more cleanly. Remember to keep track of how you resolved these issues in the past; you will find that your work gets duplicated, especially issues like removing a sign off, adding in headings, etc.

Prompt Engineering Is Not a Clinical Responsibility

All this prompt engineering seems like a lot of work for a clinician. How are clinicians supposed to save time creating documentation when they must write, test, and edit every assistant they use? Then, you get updates you didn’t plan for. Platforms can change which model your assistant is using without telling you, making all your assistants behave differently and leaving you right back at square one. Furthermore, each generation brings its own set of skills with it.

For example, in times past, it was very important you don’t tell the assistant what not to do “Don’t add in the patient’s address.” which could leave you frustrated and stuck trying to find another way to phrase that. But nowadays assistants have gotten over that issue.

So who’s meant to create, maintain and keep up with all of these assistants? At Medwrite, we provide the tools to develop your assistants and offer expert guidance from clinical prompt engineers while staying current with the latest models. But any hospital can and should have a dedicated prompt engineer. The use of AI Tools is only going to continue to grow over the coming years, and having a dedicated person on staff to react to all your clinicians' needs is a must.

‍