ChatGPT in the Doctor's Office

Physicians have long been complaining about Electronic Health Records (EHRs), about there being so much information that they can't keep track of it all, so many things they need to do that there's just too many clicks, too many functions, too much to document. As their patients’ care becomes more complex, this is only getting worse. Physicians report high rates of burnout, which they largely attribute to too many bureaucratic tasks. So how can AI help primary care physicians?

The first task is medical digital dictation: AI can take dictation, on the fly as the physician speaks, but can also do so from saved recordings as hospital transcriptionists have done in the past. Just as transcriptionists needed training to become medical transcriptionists, AI models need specific training in medical terminology.  

With the advent of natural language processing, we have made a leap forward: what if not only could AI take medical dictation but actually summarize the visit in the first place?  What if AI could also complete all the other tasks resulting from the patient visit? Essentially: act more like a real physician’s assistant rather than merely a medical transcriptionist?

Let's talk about an example: Let's say a patient goes to the doctor’s office because they’ve got a painful knee.  How might this work? 

Today, the patient walks into the exam room and the doctor asks some questions: what's the problem, how when it started, and so on.  They will do a physical exam: look at swelling, mobility, querying pain.  Throughout the exam, they’ll likely be telling the patient what they are seeing.  They will explain what they think the problem is, and what they're going to do about it; for example, that they're going to refer them to a specialist or to a physiotherapist or they're going to prescribe pain medication.

Today the physician will, at the end of the visit, have to document the referral, write the prescription, and write up a note documenting the visit as well as bill the appointment.  None of those tasks are difficult, and all are important both for the long term understanding of the patient’s physical state, for ensuring that they get the care and treatment they need, and that the physician gets paid for their work.

What if AI could do all those administrative tasks for the physician? What if AI could draft the documentation? What if AI could draft the referral, the prescription and find the billing codes and so on so all the physician has to do is review them, confirm it and they're done?  The only difference in the patient’s experience would be that the visit would start with a variation of the question: “Do you mind if this visit is recorded?”  But the difference for the physician is enormous.  

It is a great example of computers doing what computers are good at and humans doing what humans are good at.

What risk do we need to consider? Mainly, that the generator doesn't really comprehend the context, so the physician must review what is generated quite carefully.  At Tali, our trial of Natural Language Process (NLP) summarization models to create SOAP notes for a patient with a history of cancer and hypertension resulted in notes that were quite nicely written, but unfortunately included details that were invented.   

We repeated the experiment with ChatGPT itself and got a similar result – not exactly the same note of course, but a note that included some specific measurements of vitals, even though the prompt had not included any such details.  As we talked about a few weeks ago, the AI is just choosing each word in the sequence based on the probability that it is suitable, not that it is factual, correct or logical.

There are studies underway on how to use ChatGPT even more broadly: not just to generate documentation, but to use it for clinical reasoning and decision-making.  This would be a quantum leap forward, assuming the challenges of accuracy and correctness that have been highlighted to date can be addressed.  

The common thread in Tali’s implementation, a feature we call Ambient Scribe, is that the physician is very much in control: whether the physician is dictating medical notes directly, or using Tali to generate a SOAP note, the onus is still on them to review the content and ensure correctness.  Our job is just to deliver the best possible draft, and provide the best possible tools and workflow for amending the note once drafted.  This is in addition to our existing features of Medical Scribe, Medical Search and EHR Assistant. 

The potential for NLP to take tasks off of the shoulders of physicians is enormous and we believe that by doing so we will make the physicians job easier and reduce the burnout that they're experiencing today. Let computers do what computers are good at, so that humans can do what humans are good at.

Read more of our blogs here.