AI in Veterinary Medicine: Use Cases for the Hospital
Artificial intelligence is awfully exciting, and it's applications in vet med abound. There are lots of ways you can start using it in practice today.
I’m a fairly impatient guy. I’m working on it, sort of.
However, it’s been emphasized to me that there’s a lot of value in revisiting and rehashing some older concepts related to artificial intelligence. Just because I’m all about the latest and greatest updates in generative artificial intelligence doesn’t mean everyone’s caught up yet. In fact, if they were, I’d have less to write about.
For the next few posts on Substack, I’m going to be reviewing older content and updating it. The articles will look similar to longtime readers, but I’ll do my best to add new insights and research to each one.
And, frankly, while I’m still an intolerant-of-mediocrity veterinarian, practice owner, and general clinician, I’d like to think I’ve grown and learned new things over the past year or so —about artificial intelligence as well as about myself.
This first article, I hope, will serve as a review of some of the applications of AI in veterinary medicine these days.
Finally, I gave a lecture to a rather esteemed group of veterinary professionals a few weeks ago and it went very poorly. How poorly? “I asked if I could write apology letters to all of the attendees” poorly. I relied on demonstrating the technology and the technology fell out from under me. My “unusual activity” resulted in me being locked out of ChatGPT for a few hours.
If not for the kindness of another presenter, who helpfully asked questions and engaged in debate, I would’ve been left to fall flat on my face. There’s no small amount of shame that comes from such an experience, but the moderator and head of the group graciously offered me a chance to write an article that could be sent to them to support my disastrous performance.1 This will serve as that piece as well, as it contains much of what was supposed to be demonstrated in the lecture.
Real Veterinary World Applications of Large Language Models
Original: Posted April 26th, 2023.
Large language models are a type of generative artificial intelligence capable of general-purpose language generation. There are other types of AI and other types of generative AI, but I mostly confine my work to large-language models (LLMs).
Two peculiar phrases have entered my lexicon because I’ve learned about AI: “use case” is the first, and “leverage” in its verb form is the second. I really don’t like the latter.
The “use cases” of artificial intelligence, particularly large-language models, are numerous. I and others have called them “pluripotent,” a nod to the word used to describe cells that have no fixed developmental potential. That is to say, they can grow into a variety of different cell types from a single original cell. Such is the way it is with the use cases of generative artificial intelligence. So adaptable and versatile is the power of language processing that there are more applications than can be reasonably named, but I will list a few below.
Content Creation
Not everyone is creatively inclined, and not everyone needs to be. While it’s a worthy skill to practice and refine, it’s not something everyone operating a veterinary hospital has time to pursue.
So say I’m an impatient, irascible practice owner who needs to update his hospital’s social media pages more often.2 The way to do this could be to dedicate thirty minutes a day to creative pursuits needed to manage the social media pages, but anyone who’s run or even been inside of a veterinary hospital knows that empty half-hour blocks are hard to come by and even harder to protect.
I can ask ChatGPT a series of questions to help me come up with some general-purpose content.
Prompt: Please help me create a social media calendar for the next seven days that I can use in my veterinary hospital. The content should be wholesome, friendly, and welcoming. Please include ideas for each day and a suggested caption.
Response:
Here's a wholesome, friendly, and welcoming social media calendar for the next seven days for your veterinary hospital:
Day 1: Meet the Team Monday
Post: Share a photo of a staff member with a pet they have helped.
Caption: "Meet Dr. Sarah! 🌟 Her passion for animals is what makes her a cherished part of our team. Here she is with Milo, who just had his annual checkup! #TeamSpotlight #VetLife"
Day 2: Tip Tuesday
Post: Share a pet care tip, like the importance of regular grooming.
Caption: "Did you know? Regular grooming keeps your pet looking great and helps spot health issues before they become serious. 🐾 #PetCareTips #HealthyPets"
Day 3: Wellness Wednesday
Post: Post a picture of a pet during their wellness exam.
Caption: "Wellness Wednesdays keep us busy! Here’s a sneak peek at Benny's routine exam. Keeping pets healthy one checkup at a time! 🐶💖 #PetHealth #WellnessWednesday"
Day 4: Throwback Thursday
Post: Share a before and after picture of a pet that has made a significant recovery.
Caption: "Throwback to when little Zoe came in with a limp, and now she’s running around like nothing happened! #TransformationThursday #PetRecovery"
Day 5: Fun Fact Friday
Post: Share an interesting animal fact.
Caption: "Fun Fact Friday: Did you know cats can make over 100 different sounds? Meanwhile, dogs can only make about 10! 🐱🐶 #AnimalFacts #FunFactFriday"
Day 6: Shoutout Saturday
Post: Highlight a pet and their owner who recently visited.
Caption: "Shoutout to Max and his owner, Lisa, for being so brave during Max's vaccination appointment today! 🌟 #PetOfTheDay #ShoutoutSaturday"
Day 7: Sunday Snuggles
Post: Share a photo of a cozy pet in a relaxing setting at the clinic.
Caption: "Sunday Snuggles: Nothing beats a calm, cuddly day. Here’s Charlie enjoying a quiet moment in our waiting room. 💤 #RelaxingSunday #CozyPets"
Each post aims to engage your audience with a mix of educational content, behind-the-scenes glimpses, and community-focused highlights. Make sure to adjust the captions to fit your specific tone and include relevant hashtags to increase visibility!
Holy smokes, that’s way better than it used to be. This is a detailed, useful outline of a social media approach that could be used long-term. Now instead of a job that would’ve taken the practice owner or a senior employee a lot of time, it’s mostly done and can be filled in fairly easily.
And this technology can help anyone get better. Studies have shown that using ChatGPT can improve the quality of output for top-performing employees by 17% and lower-performing employees by 43%. And it’ll reduce performance equity gaps stemming from demographics and education as well.
Treating Executive Dysfunction and Decision Fatigue
This one might be more helpful to my fellow ADHD-suffering colleagues than to other folks, but it’s been awfully helpful to me. Most people experience executive dysfunction from time to time, I’ve read, and it doesn’t even have to happen 912 times each day to be frustrating. Say you’ve got an email to write and you’ve been putting it off for much longer than it would’ve taken, and the task festers on your To-Do list, taunting you with its simplicity and promising absolutely zero dopamine for its completion. Ask ChatGPT for help!
I prompted ChatGPT for help with a notice of an outstanding balance to a cat owner. If I’d given the LLM a specific owner’s name, hospital name, etc it would’ve included it in the reply. See below:
Dear [Owner's Name],
We hope this message finds you and your furry friend well. We wanted to reach out to you today regarding an overdue balance on your cat's recent visit to our veterinary hospital.
We understand that things can get busy and expenses can add up, but we kindly ask that you take a moment to review your account and make payment as soon as possible. This will help us ensure that we can continue to provide high-quality care to all of our patients.
If you have any questions or concerns about your bill, please do not hesitate to contact us. Our team is always available to assist you.
Thank you for your prompt attention to this matter. We appreciate your business and look forward to continuing to provide exceptional care to your cat.
Sincerely,
[Your Name]
[Your Veterinary Hospital]
Oh, alright, you think that one is too nice? Does it not send a strong enough message? Fine, let’s request a sterner tone and an emphatic message.3
Dear [Cat Owner's Name],
We regret to inform you that your account with our veterinary hospital is currently past due. We have made repeated attempts to contact you regarding this matter, but have yet to receive payment for the services rendered to your pet.
We understand that unforeseen circumstances can sometimes arise, but it is important to honor your financial obligations in a timely manner. Failure to do so not only affects our ability to provide high-quality care to all of our patients but also impacts the financial stability of our business.
We urge you to rectify this situation immediately by settling your outstanding balance. If you are experiencing financial difficulties, we are willing to work with you to develop a payment plan that accommodates your needs.
Please note that failure to resolve this matter in a timely manner may result in additional fees and/or suspension of services.
Thank you for your prompt attention to this matter.
Sincerely,
[Your Name]
[Name of Veterinary Hospital]
There we go! That’ll get the point across!
Perhaps more importantly, this lifts the emotional burden of an uncomfortable, delicate email from the person who’d otherwise have to write this. Nobody became a veterinarian or technician because they love writing collection letters, and this frees up that much bandwidth to do literally anything else.
This should not serve as a replacement for a person’s judgment or common sense, and I’d never send a message like that without reviewing and editing it. I should note that there are ways to refine the writing delivered by ChatGPT as well with more sophisticated prompting. And while I proclaim intolerance of mediocrity (see above), I have to admit that there are some everyday messages that can be delivered without the headiness and weight of the author’s quill.
Decision fatigue isn’t like the fatigue that comes from physical exercise, it requires similar levels of effort whether you’ve chosen antibiotics or pizza toppings. And freeing up the bandwidth and reducing the decision fatigue of folks whose decision quality can, sometimes, be the difference between life and death seems like a prudent function use case.
Translating “Doc” to “Human”
There are floating bookshelves that wrap around the walls of my hospital’s lobby, about eight feet off the ground. They hold a few hundred volumes I’ve read over the past few years. I give the books away to clients and friends when recommendations come up in conversation, always given and never lent. When the shelves were first installed, the woman who is now my wife teased me, “Oh, so you’re holding all this knowledge out of reach and over everyone’s head?” I insisted we find a better metaphor immediately, but the joke stuck with me.
Some of my loved ones will gently tease me for not always “speaking Human,” given my occasional-but-apparently-unavoidable-and-still-regrettable tendency to talk like everyone went to veterinary school.4
The use of overly technical jargon is pervasive in our industry. Heck, my locum surgeon will use phrases like “negative radioulnar incongruity” with me like I’m supposed to remember the definition of a case I’ve seen maybe once in the last decade. I’ll dumbly scratch my head and ask that he remind me what the term means.
ChatGPT can help with this problem too! It can “translate” or even write client communications that can clearly convey a message to clients or others without heavy use of medical jargon.
Prompt: As it pertains to dogs, what does the term “negative radioulnar incongruity” mean?
Response: Negative radioulnar incongruity in dogs refers to a condition where there is a misalignment between the radius and the ulna, the two long bones in the forearm. Normally, these bones should fit together smoothly at their ends to form the elbow joint and allow for proper limb function. In cases of negative radioulnar incongruity, the ulna is shorter relative to the radius. This discrepancy causes the radius to bear more weight and can lead to improper joint alignment.
This misalignment often results in joint instability, pain, and an increased risk of osteoarthritis. It can also lead to elbow dysplasia, which is a common cause of lameness in dogs. Treatment options typically include medical management to control pain and inflammation, or surgical interventions to correct the bone alignment, depending on the severity of the condition and the overall health of the dog.5
Prompt: Can you explain this in terms accessible to someone with an eighth grade education?
Response: Sure! Negative radioulnar incongruity in dogs is when the two long bones in the forearm, called the radius and the ulna, don't match up right at the elbow joint. Think of it like two puzzle pieces that don't fit well together. The ulna is shorter than it should be compared to the radius. This makes the elbow joint not work as smoothly as it should, which can be painful for the dog and might make it hard for them to move their leg properly.
Because the bones aren't aligned right, the elbow can get damaged over time. This might lead to arthritis, which is when the joint becomes swollen and sore. The main ways to help a dog with this problem are giving them medicine to reduce pain and swelling, or sometimes surgery to fix the bones so they match up better.
Medicine isn’t done if it’s not communicated. Communication is probably the most ubiquitous aspect of the job for which we receive the least formal training. Seriously, if they spent as much time teaching us communication as they did teaching us about White Muscle Disease, we’d all be better off for it.
And these large-language models present an opportunity to help us communicate with clients on any level, in dozens of languages. The LLMs can shift tone, tense, meaning with ease. 19% of American adults are at or below the PIAAC Level 1 competency, the level at which someone can read brief texts on familiar topics to extract a single piece of information. That’s one in five of our clients who would struggle to read even the most basic material.
Ignoring the fact that we aren’t very good communicators, let us focus on the fact that one in five of our clients would struggle to decipher the information in their own pet’s medical records. That’s all the reason I need to use a free tool to try to get better at it, I think it should be reason enough for you too.
A Change of Tone
Say, for example, that you’re a person whose personality borders on the intense. Maybe you’ve crossed that border a little bit. Maybe you have crossed that border with the momentum and enthusiasm of an Olympic sprinter breaking the tape. You’re direct —even impassioned, animated —even ferocious, assertive —even intimidating. You barely notice it, but it can, at times, rub some folks the wrong way as the intensity of your personality scorches the earth.6
On the other hand, maybe you have the inverse of the former personality quirk, maybe you have a hard time expressing yourself in an adequately assertive fashion. You’re trying so hard to be nice and gentle with people, a reflection of kindness and character, that you don’t always manage to speak up as effectively as you’d hoped.
You can enter your words into ChatGPT and ask it to help you change the tone of a piece. It's pretty amazing that an artificially intelligent robot trained on huge chunks of mankind’s knowledge can identify tone and help you alter that of your work. It’s one thing for an information processor to be more knowledgeable than a human, but that a language processor can be nicer is, to me, astonishing.
Here are two email examples:
Version One:
Subject: Urgent Reminder - Project Deadline
Dear Team,
I'm writing to remind you that the project deadline is approaching quickly, and we still have a lot of work to do. I need everyone to stay focused and commit to whatever is necessary to meet the deadline. If we miss the deadline, we will lose the client, which would be a huge setback for the company.
Please prioritize this project over all other tasks, and let's work together to get it done on time.
Regards,
[Your Name]
Version Two:
Subject: Friendly Reminder - Project Deadline
Hello Team,
I hope this email finds you well. I wanted to check in with everyone and remind you that our project deadline is fast approaching. I know we've all been working hard, and I appreciate your dedication to this project.
I also wanted to remind you that the success of this project is crucial for our company, and I'm confident that together, we can achieve our goals. I understand that we all have other responsibilities, but if you could prioritize this project over other tasks, it would be greatly appreciated.
Please let me know if you need any support or if there's anything we can do to help you meet your deadlines. Let's work together to deliver a successful project on time!
Warm regards,
[Your Name]
We see the variety, the change in tone. It’s powerful to see a machine deliver two dramatically different linguistic tones while maintaining the same message. Any leader knows that a lot of good or bad can be done by changing the delivery of such a message. Still, it can be hard to set aside bad emotions in communication and let our better judgment, cooler heads, and kinder words prevail.
This helps us to do our jobs better. It’s not a replacement, but an augmentation.
Medical Records
My favorite use case, by far, is the use of dictation technology to ambiently audio record and transcribe the description of a physical examination spoken aloud during an appointment and translate it into a written record in SOAP format complete with client-friendly summary.
I don’t know anyone who became a veterinarian to pursue their lifelong love of writing medical records.
There are a number of makes and models of these AI scribes available on the market, but they are powerful enhancers of doctor productivity and satisfaction.
The numbers on this are dramatic.
Completing medical records and billing records is so necessary and time-consuming that it’s often completed after-hours during what some physicians call “pajama time” and contributes significantly to physician burnout.
The federal government subsidizes electronic medical record adoption, but physicians are concerned about poor EMR usability, time-consuming data entry, reduced patient care time, inability to exchange health information, and templated notes.
80% of medical progress notes are copied and pasted, propagating errors.
Completing EMR accounted for almost 17 minutes per patient encounter.
When was the last time your physician spent 17 minutes talking to you?
Completing EMR accounted for at least 17% of nurses workload.
Relieving the burden of medical records dramatically reduces burnout in physicians.
Using an AI scribe to enhance record-keeping resulted in:
81% of patients reported the doctor spending less time looking at a screen during the visit.
Reduced time spent on EMR
Physician-reported “transformational impact” on their care.
We don’t have a productivity crisis, we have a bandwidth crisis. And I can all but solve it overnight with the widespread application of AI scribes.
Limitations
If I drive my car from my house to the airport, it will do a fabulous job of getting me there faster and making it far easier to transport family and luggage all that distance. However, if I try to drive it atop the Delaware River without the bridge, or if I try to pull up to the ticket counter still behind the wheel, or park somewhere I’m not allowed by the copious signage, I will find myself disappointed in the outcome.
An AI communications expert said it best, “Your car is not a boat.”
The jagged edge of its performance, as well as its probabilistic language processing and generation, means that it’s possible to make unexpected errors. These errors look and are fundamentally different than the “File Not Found” or “Limited Connectivity” messages with which we are all by now painfully familiar. The sneaky part of the errors from generative artificial intelligence is that they won’t announce themselves as errors. You have to notice for yourself.
I’ll put it this way: my banking software calculates how much I have in my account. Machine learning would predict it. Machine learning would be as undesirable as banking software used to tell customers their account balance. It’s quite useful in banking software for predicting which purchases that might be considered “unusual” for security measures.
Predictions are occasionally incorrect, but calculations, as done by traditional deterministic software, are basically never incorrect. There are times when predictions are acceptable and times when they are not. It is time that we began to know the difference.
A specifically trained AI model outperformed physicians in creating a list of ten differentials for cases taken from the New England Journal of Medicine. But every doctor knows that a differential diagnosis is not a diagnosis, and there is, at times, a vast chasm separating the two.
All of this is to say that large-language models are a tool that can augment the medical practitioner but not replace them. I suspect that my status as a clinician makes me less bothered by the pressures of outside error. If I consult a specialist who gives bad advice that leads me to make a bad decision, the fault is mine. If I consult a dated textbook that leads me to make a bad decision, the fault is mine. If I consult a large-language model that leads me to make a bad decision, the fault is mine.
A decade of hands-on veterinary practice has left me with a pretty comfortable acceptance of the fact that the buck stops with me. I like it that way. The pressure that comes with responsibility and accountability galvanizes focus and channels intensity. I don’t look for excuses, I look for answers.
As veterinarians, we have the responsibility to understand each and each drug, tool, and technology we use in the care of our patients. We must know how these things work and how to use them safely and correctly. That level of responsibility doesn’t change for the scalpel blade, the stethoscope, or the large-language model.
Conclusion
I don’t like repeating myself, but I believe there’s value in reviewing older concepts and bringing them once again to the forefront of the mind. It’s why we take notes, attend continuing education courses, and constantly study our craft.
In reiterating these ideas, I am not abandoning my commitment to continuous learning but rather taking the time to make an effort at continuous teaching. Not everyone has years of experience studying and researching LLMs. Not everyone has the ability to implement large-language models in their hospital. Not everyone has a newborn who keeps them up all night so they can burn through audiobooks.7
I believe these concepts, this technology, and these applications can help my colleagues in the same dramatic ways they have helped me. My Substack is free and I don’t hide content behind a paywall. I’m a scientist at heart, I think sharing knowledge makes the world better. Possessing knowledge doesn’t necessarily impart ownership of it, and I deeply resent those who would claim the work of others as their own as well as those who would count themselves as gatekeepers to such knowledge, keeping it behind the high cost of conferences or consulting fees.
So I try to cite my sources, credit the original authors, and share what I know as freely as I can. I try to be generous with my time and expertise, though I admit sometimes I fall short of my own ideals. I’m intolerant of those who would waste time or who would sell their skill in self-promotion as genuine subject matter expertise.8 Such intolerance, as well as the aforementioned9 intensity, inevitably bothers some people once in a while. I’m okay with that; at least, I’m trying to be.
What I do as a veterinarian is important. And, sometimes, part of doing something very well is pointing out what's going wrong.
The next part is figuring out how to get better. And that, like technological advancement, is always a work in progress.
This article should’ve been written weeks ago. But newborns are hard and take up a remarkable amount of bandwidth. He’s awfully cute though.
Based on a true story.
I almost never have this problem.
There are two types of veterinarians: the ones who acknowledge that they do this and the ones who deny that they do this.
Oh right! Just needed a little jog of the memory.
This is entirely hypothetical.
Some parents play their kids Mozart. My kid gets Ethan Mollick. I’ll let you know how that turns out.
Knowledge and skill in self-promotion is its own expertise, which I don’t mind and will gratefully purchase when the need arises.
And hypothetical.