See the Full Issue
June 2022
Next-Gen Diagnostics
Mike Petty, DVM, reviews three different AI and algorithm-assisted veterinary diagnostics systems: a system that reads radiographs, one that diagnoses pain, and another that diagnoses internal medicine problems.
A Review of 3 Artificial Intelligence and Algorithm-Assisted Systems
Is there any one of us who didn’t know, suspect, or even fear that someday artificial intelligence (AI) would play a role in veterinary medicine? Like it or not, we are at that point. After having worked with three different systems, I am happy to report that in my opinion AI is not here to replace us, only to assist us.
I have reviewed three different systems for the purpose of this article: a system that reads radiographs, one that diagnoses pain, and another that diagnoses internal medicine problems.
There are many more in the works, and we can expect AI assistance in areas such as anesthesia and rate calculations as well as oncology. In the discussion that follows, I am going to list what I liked and didn’t like about each of the three systems.
Radiology
You may already be using the system out there for reading radiographs by SignalPET. Basically, how it works is that you choose the suspect areas as you normally would, upload the radiographs and then receive a report. The report lists all areas looked at and they are diagnosed as normal or abnormal.
Additionally, each abnormal result then generically describes why it was listed as abnormal and gives a few differentials to consider and maybe some suggested follow-up diagnostics such as an ultrasound. See the boxed area below for an example of an abnormal finding from the system of Urinary Bladder Calculi.
Urinary Bladder Calculi: |
|
| An abnormal test indicates that an abnormal/mineral opacity has been detected over the outline of the bladder. The most likely diagnosis being urolithiasis, other causes can be ruled out with additional diagnostics such as additional radiographic images, contrast study, abdominal ultrasound, etc. DDX: Urolithiasis, other | |
As you can see by the description, all bases are covered with reasonable additional diagnostics suggested.
What I liked about SignalPET
I have been reading radiographs for over 40 years, but I still have the tunnel vision of looking at the main area of interest and using it for confirmation of what I suspected. If I am in a hurry, I sometimes do not go back and look at every structure at the edge of the radiographic field and important observations are sometimes missed. This program looks at absolutely everything, and I am humbled to say that a few things like stifle effusion were picked up even though there was no stifle pain, the main source being the hips and pelvis and where I focused my observation. This program gives you a second set of eyes.
What I disliked about SignalPET
Sometimes it is just wrong. In the above example of the Urinary Bladder Calculi, it was actually a piece of radiopaque stool that was superimposed on the bladder. It was obvious in the other view that this same density was nowhere near the bladder.
The other issue is that it does not diagnose everything. The user is provided a list of things the AI looks for, but it is up to us to look for other issues. In other words, it could be very easy to become complacent and not make your own observations.
Conclusion
Maybe someday the AI will be able to compare one film with the other. And when talking to the software developers, they assure me, and I believe them, that it will only get better with time, both in finding mistakes it might make and in diagnosing additional problems.
Pain Diagnostics
A dog receiving laser therapy during real-time monitoring of pain signals with the PainTrace device. (Photo courtesy of BioTraceIT)
In the PainTrace system, produced by BioTraceIT, an animal is monitored through a variety of joint manipulations, such as flexion and extension of a shoulder or hip, while connected to a wearable device by electrodes.
The system has the ability to monitor abnormal nociceptor activity and report it on a screen with a real-time visual. This actually works two ways: it shows pain both in a resting state and during movement or manipulation.
For example: If an animal has pain issues, and it is connected to PainTrace, it may show a chronic level of pain, even at a resting state. This does not tell us where the pain is. In order to do that, each joint needs to be manipulated to register a reading on the device and diagnose it as pain, and it also provides the degree of pain based on the amplitude of the response.
What I like about PainTrace
It is very good at detecting pain. I have even used it on anesthetized animals and gotten positive pain readings. On awake animals, it has found areas of pain that I knew existed. Unlike a pain metrology survey, it cannot be fooled by an owner that doesn’t want to answer the survey correctly, or a veterinarian that does not have the skill set (or patience) to do and interpret a complete pain exam.
What I don’t like about PainTrace
There is a learning curve to attach and use the device properly. There is also a minor delay, one to two seconds, between a painful manipulation and the readout on the iPad, and the user must have the patience to move a suspected joint.
Conclusion
I see this device as a very helpful tool not only for the hard-to-diagnose cases but also for the stoic animals that do not show outward signs of pain as well as for the practitioner who is not confident in their pain diagnostic skills. Additionally, clients appreciate the visual ability to locate painful anatomy and track treatment effectiveness overtime; it is a great communication tool.
Internal Medicine Diagnostics
Gekko 7: A screenshot of the GekkoVet system, showing a list of differentials for an unspayed female golden retriever with polyuria and leukocytosis. Photo courtesy of Michael Petty, DVM
GekkoVet is probably the most advanced player in the realm of internal medicine diagnostics. In full disclosure, I was a beta tester two years ago and made suggestions to improve the system.
Basically, you can either enter findings in a generic mode (male neutered cat, for example) or set up a case study by entering breed, age, geographic location and other parameters. Then you enter your findings, which can include client history, physical exam findings, lab results, and radiographic findings. The AI then searches a database of tens of thousands of pages of articles and lists diagnoses in order of probability. It does this in about a second.
You can then look at each possible diagnosis and the software will list rule-in/rule-out testing. Once you have chosen a diagnosis, it will then give you up-to-date treatments and even includes a drug formulary.
What I like about GekkoVet
We all have “those” cases. When you have practiced as long as I have, there are not as many as there are for new graduates. Even so, sometimes I am surprised by the list of possible diagnoses, which is the problem of practicing so long. Sometimes we start to get tunnel vision based on certain sets of signs and history and forget about some of the outliers that need to be on the list should our initial thoughts be wrong.
Additionally, “casting a wide net” and running too many tests because you know “something ain’t right” can quickly alienate a client because of time and money spent. Having a defined list of rule-outs from a software program like this to present to the owner makes them part of the diagnostic team, and you are less likely to chase a diagnostic red herring.
What I don’t like about GekkoVet
Too much information put in can give possible diagnoses in the hundreds. You have to pick and choose what you decide is important. For example, if an animal is severely dehydrated, then adding tachycardia and poor capillary refill is not necessary, but the program doesn’t know that and will start looking at tachycardia and poor capillary refill as primary issues and increase the diagnostic list. In other words, you need to trust and wade through your own clinical observations as well as decide what history from the owner is pertinent in order to narrow your list down.
Conclusion
The judicious use of diagnosis assistance can not only help shape our diagnostic acumen for future cases but also relieve the stress associated with those difficult case presentations while educating us and making us better diagnosticians.
Summary
I don’t foresee a future wherein the veterinarian is going to be replaced. The subtleties of diagnostics vis-à-vis taking a proper history, doing a thorough exam, ordering the right lab work or radiographs, and so on are too complex for any AI to make that decision on its own.
I do see a future in which our workload can be reduced: not stressing about hard cases, streamlining the rule-in/rule-out process, and spending time we might not have for looking up the latest therapies will become a thing of the past. It is up to the individual practitioner to decide when and how much to embrace AI, but the future is here and we are going to have to change with the times sooner or later.
Michael C. Petty, DVM, CCRT, CVPP, DAAPM, is in private practice in Canton, Michigan. He is a frequent national and international lecturer on topics related to pain management. Petty offers commentary on each Pain Case of the Month (and occasionally writes one himself). He was also a member of the task force for the 2015 AAHA/AAFP Pain Management Guidelines for Dogs and Cats. |