Welcome back dear readers to another post of random health technology insights! Today’s topic sounds well a-bit niche…. Yes it is but hear me out there is a reason for this randomness, Copilot Agents
Picture this, it’s 3 AM and I’m jolted awake by the sound of my toddler crying again… After the obvious quick temperature check, I realise the poor (not so) little thing has a fever. TBF he was coughing the night before and generally wasn’t in great form so should have seen it coming really!
So, I do what any sleep-deprived parent would do I administer some paracetamol and try to soothe them back to sleep.
The next day, he’s no better and we arrange a visit to the doctors, after the standard 8am phone line queue we get an appointment and they confirm he has got a chest infection, and we get some antibiotics.
So after a couple more broken nights sleep and some very early morning wake ups until the antibiotics kick in! I am sitting there bleary-eyed watching the #MSIgnite conference and clutching my coffee, trying not to drift off! It suddenly dawns on me….
We need primary care knowledge Copilot Agents
Yep random right! So we all know how capacity stretched all of our awesome healthcare workers are on the day to day. Wouldn’t it be amazing if there was a way for them to rapidly access a wealth of clinical knowledge to help guide and inform them to diagnose symptoms and suggest potential treatments all from a trusted online information source?
Why yes it would! And thus, the idea for the primary care knowledge agent was born!
So by now the antibiotics were working and my son was better, so I brewed yet another cup of coffee and set myself a challenge, how fast could I make something using this awesome new agentic Copilot technology we now live in post #MSignite, and guess what the answer is pretty damn fast…. like super damn fast..
So dear readers, sit back grab your calpol and coffee and let’s dive into the process of how to build a primary care knowledge agent!
A coffee fulled vision… let’s enhance diagnosis with Copilot Agents
Now imagine a world where clinicians have instant access to the latest clinical guidelines, drug information, and treatment options, all at their fingertips. This is the coffee fuelled vision behind the primary care knowledge agent.
By leveraging the awesomeness of autonomous agents and the power of generative AI, we could support care providers in their diagnostic process by using AI to delve and search across the wealth of information available and indicating matches to provided symptoms and detailing treatment options, ultimately giving clinicians time back to focus their patients.
Creating a clinical knowledge agent might sound like a scary and daunting task, but with coffee & Microsoft Copilot technology, it’s surprisingly straightforward which leads me onto the next point.
A Clinical Safety Word of Warning
With great power comes great responsibility! The responsible use of AI technology has some considerations here, whatever the engine comes back with needs to be taken with a pinch of salt.
AI Hallucinates, it is after all a probability engine, returning what it thinks is the likely response based on the inputs from the user, as such should not be used as a pure diagnostic tool, a person should always be making the final decision.
So before anyone says, I am absolutely not saying Copilot is clinically validated in anyway and the outputs should absolutely be reviewed by a clinically trained person before anything is prescribed to a patient.
So with warnings heeded, let’s dig into how simple this process is!
Getting started let’s get us some educational data
Any AI Agent is only as good as it’s data source, now having been in and around front line healthcare, I have often seen and heard clinicians referring to things called the “CKS”, “BNF”… “National Formulary” and “NICE Guidance”, even my mother had one of these BNF’s lying about the house as a child
So in the UK, NICE is the central body who collates all this information which is made available for NHS clinical staff to use, the information is kept up to date and is one of the main sources when it comes to conditions, medications, doses etc.
In terms of clinical knowledge specifically this is all housed in the CKS, Clinical Knowledge Summaries and this is available as a public website, you can see where this is going…
A Licensing word of warning
So, another warning, there’s lots of warnings in this AI world…
As per the terms of use for the CKS I am not in anyway affiliated with NICE, or operating as a clinician and I am absolutely using the data in a purely educational purpose and it is not being used on behalf of or for the benefit of any company, organisation or business. I am also not using the data to self diagnose or as a substitute for a consultation with a qualified clinician!
How to build an educational Clinical Knowledge Copilot Agent
So with warnings heeded off to build our world changing agent and it’s pretty simple, first you need to head to: https://copilotstudio.microsoft.com
Then simply login and click “create new agent” you will be then presented with the following form, to create the agent you populate the values as follows.
Name: – Simply provide a descriptive name for your new Agent, lets go with “Clinical Knowledge Agent”
Description: – This field is essentially an internal description value for the agent, fill it out with a few simple descriptive words.
Instructions: – This is where the magic happens, this field is essentially the prompt to Copilot which explains to the agent how it should operate and handle the data and responses to the users, it is in effect how you prompt engineer the agent itself, as opposed to a specific prompt in the conversation.
Knowledge:- This is where you enter the source of knowledge for the agents, in this case we are using the CKS website, but equally it could be a set of internal documents, intranet site or even more complex data structures such as a database or another system via APIs etc
With that completed, click “Create”
And the agent is born, almost in 007 seconds…..
And we are done the agent is ready, nearly.. Ahh Mr Bond we have been expecting you.
Now the final step is to switch on generative AI in the orchestration section, which effectively enables the generative engine to be able to manage the conversation and responses as opposed to following a prescribed response based on specific topics
Final step, click Publish …. see told you it was quick!
So what does it give us? a pretty damn good reference agent!
Well let’s test this out, shall we, so far it’s taken us 10 minutes to build this agent, surely it can’t be able to provide something clinically relevant that quick?
Enter our patient to our virtual clinic room, they are there in a purely educational basis and they would like to discuss an imaginary rash on their leg. They provide the details of the symptoms and the current state of their feeling.
The primary care clinician has a suspicion but would just like to quickly find some reference material to support their thinking, they enter the details of what the patient has provided and low and behold the agent returns what it could be, along with a link to the reference article from where it has found the information
Sounds like what the clinician was thinking, so they have an idea of the cause but what can they do about it, a quick question is posed to the agent of what would be the best course of treatment?
And sure enough the agent has that covered also, in this case we have a course of antibiotics, and also the frequency of the course, and an a couple of options in case the patient has an allergy to a particular type.
Now as any clinician knows, you can’t provide an prescription without checking for any potential interactions, well a quick question and we can also find that information
Sure enough there are some considerations for each and the agent details all of these out for educational purposes, and as any responsible AI service should be it also recommended that this treatment is reviewed by a clinical professional and it’s not a substitute for professional medical advice.
That’s awesome but not very … well autonomous #disappoint
Ok yes granted there’s not a lot of autonomy there, but you can’t have autonomy and quick!
The autonomous capabilities announced at #MSignite enables the agent to be triggered based on certain key actions, like an email received for example, or record updated in a database or even a file uploaded on a SharePoint site.
Meaning, say if a clinician had an email sent to them and the body of the email contained a reference to a drug or condition, the agent could autonomously read the email and them message the clinician to say they it has a noted a reference to a condition and they might like to consider this drug or there is a know allergy.
When you add into this workflow the ability to integrate with other systems and API’s as I have blogged about before the power of these agents really starts to boggle the mind, the potential cross knitting of information and workflows is huge.
So, there you have it 10 minutes end to end, that’s damn fast
Well if thats not enough to blow your mind then your mind is unblowable! All jokes aside (and there were a few bad ones) it just shows you how quick it is to build an agent that can provide some pretty interesting and useful results!
In the real world however and especially in a clinical setting it would never be this quick as there is a whole host of clinical saftey process to be followed.
But given the rapid advances of AI technology it wont be long till this sort of experience its readily available, I certainly hope it won’t be long!
With clinical staff spending an excessive amount of time accessing different systems of reference, pulling data out of text books and keeping it in their own heads, systems such as this make information readily available in the flow of work, it’s a major game changer! Talk about productivity enhancements
Not bad for a coffee fuelled vision, maybe I need to drink more coffee…
