Work is being done to help ease the daily workload of Alberta physicians and make the health-care system more efficient by using artificial intelligence.
Ross Mitchell, a fellow at the Alberta Machine Intelligence Institute (AMII), is doing some AI work for Alberta Health Services and the University of Alberta.
He’s created large language models — with set parametres and behind a firewall — that can make doctors’ day-to-day tasks easier.
“I’ve got a version, a little mini ChatGPT if you will, and I’m getting it to read and summarize doctors’ notes. This can pull information out of free-format texts,” Mitchell explained.
“With these new models, the ChatGPT’s capabilities, you can get them to read notes and extract information from them extremely well and very, very quickly. Faster than I can read it, it can read it and pull the information out that I’m interested in,” he added.
“It can figure out things like what the doctor’s specialty is based on the context of the note. This is a gastroenterologist. This is a neurologist. And you can ask it: ‘What are the major findings reported in this note?’ And it’ll read the whole report and tell you the major findings and leave out the minor ones.”
Mitchell sees this technology being used in a variety of ways in medicine, for things like summarizing medical reports, transcribing doctors’ appointments and even interpreting lab results.
“If you were talking to a patient in the exam room, imagine being able to say: ‘What did the radiologist say about Mrs. Smith?’ and there’s a bit of a pause and then you hear the computer tell you the summary of the radiology report because it’s gone and found it for you.”
Get weekly health news
“It would be like interacting with an intelligent assistant.”
Patients could use it as well, Mitchell says.
“Say you’ve got this lab report on your portal, imagine being able to say: ‘What does this mean?’ and it could go through and say: ‘Well, this one and this one and this one suggest you might have these things, go ask your doctor.'”
It could also save time for doctors when it comes to making patient notes.
“Instead of you having to do that at night after the patient’s gone and you’ve seen 15 patients that day and you’ve got to go through and listen to the recording and summarize everything and you type it in yourself, it might take you an hour. This could do each patient for you in a few seconds.”
Another application, Mitchell says, is finding suitable patients for new clinical trials and health research.
“We build a system to go through and read notes and summarize them and send the results to a doctor and that would help them find patients for a clinical trial for new experiments.
“Before, we would have had to hire someone to go through and read all these reports and write down what the interesting bits and pieces are. Now we can use a computer to summarize that, so we can run much larger studies much more quickly to test new things in health care.”
This model has set parametres in terms of what verified sources the AI is allowed to search, which is something models like ChatGPT don’t have. This one is also running behind a firewall.
“One problem we have now with these large language models is they’re not on a leash, they have no guardrails and they’re compelled to provide an answer, so they make stuff up. And a lot of it is just wonky, and so there’s a risk to that, especially in health care. So what we want to do is validate these models and put the guardrails on.”
Mitchell says the requests made to the model can also specify where it can search for answers.
“This is me typing into the model: ‘You’re a well-educated healthcare professional and you’re very good at extracting information from medical notes. Take the following medical note and answer the user’s questions as best you can. Your answer must come from that medical note. If you do not know the answer just reply: ‘The information is not there. I do not know what the answer is.'”
Mitchell says there’s potential to scale this up widely. He’d like to see it used across Alberta health-care settings. The technology is there, he says, it’s now just a matter of computer capacity.
“It requires a very high-end powerful, specialized computer that’s designed for these AI algorithms.”
Comments