The University of Kansas Health System is set to trial software designed to help doctors automatically generate notes from conversations with patients, a move billed as "the most significant rollout to date of generative AI in healthcare" yet.
The technology, developed by Pittsburgh, Pennsylvania startup Abridge, aims to reduce workloads for clinicians and improve care for patients. Shivdev Rao, the company's CEO and a cardiologist, told The Register doctors can spend hours writing up notes from their previous patient sessions outside their usual work schedules.
"That really adds up over time, and I think it has contributed in large part to this public health crisis that we have right now around doctors and nurses burning out and leaving the profession." Clinicians will often have to transcribe audio recordings or recall conversations from memory when writing their notes, she added.
"Just to give you a sense of what status quo can look like right now, I might see patients and then not document until the weekend, and then I'm using chicken scratch that I put on a piece of paper to remind myself what we talked about. And so it's very lossy - I'm losing a lot of the details in this document," Rao said.
Abridge's software automatically generates summaries of medical conversations using AI and natural language processing algorithms. In a short demo, The Register pretended to be a mock patient talking to Rao about suffering from shortness of breath, diabetes, and drinking three bottles of wine every week. Abridge's software was able to note down things like symptoms, medicines recommended by the doctor, and actions the clinician should follow up on in future appointments.
The code works by listening out for keywords and classifying important information. "If I said take Metoprolol twice, an entity would be Metoprolol, and then twice a day would be an attribute. And if I said by mouth, that's another attribute. And we could do the same thing with the wine example. Wine would be an entity, and an attribute would be three bottles, and other attribute every night."
"We're creating a structured data dataset; [the software is] classifying everything that I said and you said into different categories of the conversation. But then once it's classified all the information, the last piece is generative."
At this point, Rao explained Abridge uses a transformer-based model to generate a document piecing together the classified information into short sentences under various subsections describing a patient's previous history of illness, future plans or actions to take.
Generative AI is currently all the rage right now. Big Tech is desperately finding ways to incorporate these types of models into their services and a wave of startups are racing to build products on commercial APIs like OpenAI's ChatGPT. But these models often make up false information and struggle to produce accurate content, making applications in industries like healthcare or law particularly risky.
Physicians can edit the notes further, whilst patients can access them in an app. Rao likened Abridge's technology to a copilot, and was keen to emphasize that doctors remain in charge, and should check and edit the generated notes if necessary. Both patients and doctors also have access to recordings of their meetings, and can click on specific keywords to have the software play back parts of the audio when the specific word was uttered during their conversation.
"We're going all the way from the summary we put in front of users and we're tracing it back to the ground truth of the conversation. And so if I have a conversation, and I couldn't recall something happening, I can always double-check that this wasn't a hallucination. There are models in between that are making sure to not expose something that was not discussed."
Abridge's software has reportedly been trialed by over 2,000 clinicians and to help more than 200,000 patients. Now, the company is going to test it out in a real hospital starting with a small cohort of doctors.
Gregory Ator, Chief Medical Information Officer and Head and Neck Surgeon at The University of Kansas Health System, said he expected the trial to start soon.
"We need to figure out who are the ideal clinicians for this type of solution. First of all, we're going to validate that the datasets used [to train the software] are accurate and appropriate to Midwest healthcare, and deal with the type of problems that we deal with, and then we will progressively roll it out," he told The Register. ®
Will be followed soon after by SLE 15 SP 5 as org continues prep for ALP
Boffins and machines write very differently - and it's easy to tell
Brush up on your coding - more tech jobs are going to be hybrids that mix ops and software, or require AI skills
Suggest that Zuck has yet again unleashed stuff without a thought for the downsides
WWDC And makes developer-grade OS betas available to all ducking loyalists
Deadly accident said to be unavoidable
This could be a useful way to show what you're up against, or give the clueless a stick to beat you with