This appeared last week.
Best Practice will use AI assistant to draft GP patient notes by ‘listening’ to consults
Best Practice is integrating an AI assistant for note-taking.
Major GP software maker Best Practice is integrating a note-taking AI assistant to its desktop software.
The tool was built by Lyrebird Health, an Australian company founded eight months ago and will be fully integrated into the software which is used by around 25,000 doctors.
Best Practice claims it will save GPs who choose to use it between 60-90 minutes of note-writing every day.
How does it work, how was it tested and are there potential medico-legal risks?
Danielle Bancroft, the company’s chief product officer, answers AusDoc’s questions.
AusDoc: The Lyrebird tool is meant to write up clinical notes based on ‘listening’ to the consultation. How does it know what’s important to record?
Danielle Bancroft: The tool is already programmed to know what is clinical information and reject ‘chit chat’.
That process has improved as the database of doctors with access to the system has increased.
If the occasional non-clinical line creeps in, and the GP deletes it from the consultation notes, the system will ‘learn’ not to include similar data in the future.
When I first had it demonstrated to me, we were in a room with 5000 people.
We talked about all kinds of things including the weekend footy with the kids.
It managed to remove all of that.
AD: Does Lyrebird keep all the recordings from GP consults to ‘teach’ the AI based on real patients?
Ms Bancroft: No, the tool captures what is said and immediately removes any identifying patient information from the audio stream while it’s being processed, as consultation notes are generated and displayed for the doctor to review.
Each recording is only available for 24 hours after the consultation before it’s deleted.
It’s purely there for the clinician to check against and validate the notes in that period of time.
It won’t be stored forever.
The tool is really just short-cutting the manual part of typing up the initial notes.
The clinician is in control of what is actually saved.
AD: What happens during that 24-hour period?
Ms Bancroft: During a consult, all audio is transcribed in real-time on Lyrebird Health’s Australian servers.
At no point in time are audio files saved or permanently stored.
The audio stream from the consultation is completely encrypted and securely transferred to the servers.
What this means is that by the time a consult is finished, all audio has already been converted to text and there is no audio remanence of the conversation.
Even if accessed, it can’t be tied back to the individual GP or patient.
After 24 hours it is removed.
Ensuring the recording does not persist or contain personal information minimises the risk of breach or data spill.
AD: What about for the individual GP, the AI ‘learns’ what they want and don’t want in their notes?
Ms Bancroft: As an example, the system generates consultation notes with default sub-headings — symptoms, observations and the like.
If a GP removes certain subheadings they don’t use, or add new subheadings in, the notes will eventually reflect the doctor’s usual note-taking format.
But these changes are specific to that doctor.
Other GPs won’t find their consultation notes changing as a result.
The tool is an enhancement to provide a more detailed base to start from for the consult note.
It does not replace the clinician’s involvement or responsibilities.
AD: If Lyrebird notes down something incorrectly, and there are consequences, who is held responsible?
Ms Bancroft: The integration workflow ensures that draft consult note produced is checked by the clinician first before saving/writing to the database.
There is a confirmation box that prompts users to double-check and confirm the accuracy of their records before they are able to export it to Best Practice.
Lyrebird worked with medicolegal documentation experts when developing the tool.
One described the depth of information it generates as being 3-4 times greater than what they would ordinarily write.
Crucially, through work with doctors who specialise in the quality of records, there has been a significant increase in documentation quality when compared with notes manually recorded by a GP.
On average, less than 3% of the output text that Lyrebird generates is being edited.
The responsibility is still on the clinician to ensure the clinical notes are accurate prior to saving, just as it is today.
They have an opportunity to change or add to the notes before saving.
Lyrebird consulted directly with medical defence organisations when developing the tool.
AD: Given it’s based on what’s said aloud during the consultation, will GPs need to tweak their style? For example, if they’re taking blood pressure and both the GP and the patient can view the reading, will they need to announce it out loud?
Ms Bancroft: In that example, GPs will have to say the blood pressure reading [out loud].
There’s also a little bit of change required during examinations, with GPs maybe having to pronounce what they are doing, rather than just chit-chat while they check.
AD: What if a GP or patient has a strong accent?
Ms Bancroft: [We’ve done work] concentrating on different accents.
The more people who use the system the better that is going to get as well.
It also gets to know the individual clinician: how they speak, how they interact with their patients and their approaches.
AD: Will GPs need to secure patient consent specifically to use Lyrebird for note-taking?
Ms Bancroft: When GPs hit record at the start of the consultation, a prompt will ask if they have obtained patient consent and will record whether the patient has said ‘yes, just for this consult’ or ‘yes for all consults’ or ‘no’.
We are working on a workflow that will allow clinics to send out a SMS with a link so patients can be informed and consent before they get to the clinic at reception.
AD: Has a system similar to this been used anywhere else by doctors and in what context?
Ms Bancroft: Yes, one example being clinical decision support tools that process information about the patient, their history and risk factors and measures it against known information and trends to provide prompts and suggestions back to the doctor.
MIMS is another example.
That information is provided based on certain triggers and known trends but ultimately the clinician is the one who takes in the information and decides next steps.
AD: Have GPs tested the system ahead of its release?
Ms Bancroft: Yes.
We have a number of beta testers currently in the process of testing the integration of Lyrebird with Best Practice in advance of the Orchid Sp2 release in a few weeks.
The metrics we use include user experience, efficiency, accuracy and future workflow enhancement.
Best Practice clinical adviser Dr Fabrina Hossain also utilised the tool within her travel medicine clinic and focused on efficiency in terms of time saved through the day, on average 60-90 minutes per day, through not running late or staying back to flesh out and record consult notes.
More here:
We need to wait for the reports from the field to see how it actually performs in practice. Exciting times!
David.
No comments:
Post a Comment