#5391: The Software Behind Urgent Care Triage

Big buttons, emoji vitals, and a system that says "order IV" — what's actually running behind the triage screen?

Featuring
Listen
0:00
0:00
Episode Details
Episode ID
MWP-5574
Published
Duration
21:08
Audio
Direct link
Pipeline
V5.2
TTS Engine
chatterbox-regular
Script Writing Agent
DeepSeek 4.1 Flash

AI-Generated Content: This podcast is created using AI personas. Please verify any important information independently.

The screen Daniel saw at urgent care wasn't the electronic health record. It was the triage logic layer that sits on top of it — clinical decision support, or CDS. The EHR holds the chart; the CDS layer holds the protocol. The biggest name in that protocol world is Schmitt-Thompson, standardized telephone and in-person triage algorithms developed by a physician panel and updated annually, used by over a thousand organizations and covering more than ninety percent of U.S. triage calls. TriageLogic is the company most associated with commercializing them, deployed in over a thousand hospitals and urgent care centers.

The interface looks old-school on purpose. Massive text, big buttons, color coding, emoji condition indicators — all of it is cognitive load engineering. Triage is fast and high-stress, and every pixel that isn't glanceable is a liability. A face parses faster than a word; a red block reads from across the room.

Under the hood, it's a rules engine. Inputs — vitals, symptoms, age, weight, history — traverse a decision tree written by physicians and reviewed annually. "Check if the patient vomited in the last three hours" is a node. "Order IV" is a node. The IV math runs on medical calculators like Holliday-Segar for pediatric maintenance fluids, and the resulting order routes through computerized provider order entry to a physician or nurse. The tech is the sensor, the system is the protocol, the physician is the authority.

AI shows up at the edges, not in the tree. Mednition's KATE uses NLP on nurse notes to flag sepsis risk. Epic's Deterioration Index outputs a probability, not a diagnosis. Both are narrow risk scores — a prompt to look closer, not a decision. Rule-based systems dominate because they're auditable: you can trace every decision to a node. Deep learning can't offer that, and "the model said so" isn't a defense. LLMs are worse for this workload — probabilistic by design, prone to confabulating when they hit an unfamiliar input, where a rules engine defaults to the conservative branch. Demos look good because the inputs are clean. Real triage has ambiguous timelines, mistyped weights, and screaming toddlers.

Downloads

Episode Audio

Download the full episode as an MP3 file

Download MP3
Transcript (TXT)

Plain text transcript file

Transcript (PDF)

Formatted PDF with styling

#5391: The Software Behind Urgent Care Triage

Corn
Daniel's wife and son had to go to urgent care this week. Bad stomach bug, Hannah needed an IV, and while he was sitting at the triage station facing the monitor, he watched the tech work the screen.
Herman
Which is where most of us would just... look away politely.
Corn
Daniel doesn't look away politely. He takes notes. What he saw was a big-button, color-coded interface, old-school looking, massive text, and as the tech punched in vitals from the instruments, the system was talking back. Little emoji indicators for the baby's condition. Prompts like "check if the patient vomited in the last three hours." And at one point, "order IV."
Herman
So it's not just a form. It's a conversation.
Corn
That's exactly what he's asking about. He's not interested in the intake layer, the demographics and insurance part. He wants to know what the triage support system is, the thing that's making sure the tech asks the right questions and automating some of the IV ordering. He's guessing the simple UI is sitting on top of something sophisticated, that there's decision support and medical calculators doing dosage math based on size and age. He wants to know what this class of software is called, who makes it, where it's quietly deployed. And then the big one: is this mostly rule-based, in the traditional decision-tree sense, or is reasoning-based AI already running in these systems? And if so, what diagnostic model would actually be deployed for this kind of workload?
Herman
Because the honest answer is "both, but not where you'd think."
Corn
So what was Daniel actually looking at? Let's start with the category.
Herman
The category is clinical decision support, CDS. Specifically, triage support software, and it's used in urgent care and emergency departments all over the place. The important distinction is that it's not the electronic health record. The EHR is the system of record, the chart. What Daniel saw is the triage logic layer that sits on top of or alongside the EHR. It's the thing that encodes the protocol.
Corn
So the EHR holds the patient. This thing holds the procedure.
Herman
Right. And the biggest name in the protocol world is Schmitt-Thompson. Those are standardized triage algorithms, telephone and in-person, developed by a panel of physicians and updated annually. They're used by over a thousand organizations and cover more than ninety percent of U.S. triage calls. TriageLogic is the company most associated with commercializing them, and their software is deployed in over a thousand hospitals and urgent care centers.
Corn
A thousand organizations running the same decision trees.
Herman
Which is the point. Standardization. You want the tech in Jerusalem and the tech in Ohio asking the same follow-up question when a kid presents with vomiting and abdominal pain. Then there's Mednition, with a product called KATE, which is a different animal, we'll get to it. And the bigger players, Optum, Elsevier, and Vocera, now part of Stryker, for the communication and triage side.
Corn
So what's actually happening under the hood when the tech taps that big button?
Herman
Let's start with the UI, because the UI is the tell. It looks old-school on purpose. Massive text, big buttons, color coding. That's not a failure of design, it's the design. Triage is a high-stress, fast-paced environment. The tech is entering vitals while talking to a patient, sometimes a screaming toddler, sometimes a parent who's panicking. Cognitive load has to be minimized. Every pixel that isn't glanceable is a liability.
Corn
So the ugliness is a feature.
Herman
The ugliness is a feature. You want a screen where a red block means "this is urgent" from across the room. You want buttons big enough to hit with a gloved finger while you're holding a thermometer in the other hand. Daniel mentioned emojis for the baby's condition. That's the same principle. A face is faster to parse than a word.
Corn
And behind the face?
Herman
Behind the face is a rules engine. For TriageLogic, that's the Schmitt-Thompson protocols. Think of it as a decision tree. The system takes input, vitals, symptoms, age, weight, patient history, and traverses the tree. At each node, it either asks for more information or triggers an action. "Check if the patient vomited in the last three hours" is a node. "Order IV" is a node. It's classic rule-based logic, not AI.
Corn
So when the system asked the tech to check for vomiting, that wasn't a model reasoning about dehydration risk. That was a branch in a flowchart.
Herman
That was a branch in a flowchart. And the flowchart was written by physicians and reviewed annually. It's deterministic. Same inputs, same path, every time.
Corn
Which is exactly what you want when the cost of a wrong turn is a missed diagnosis.
Herman
And the IV ordering piece, that's where the medical calculators come in. IV dosages are calculated based on patient size, age, weight. For pediatric maintenance fluids, the standard is the Holliday-Segar method. It's a formula, not a guess. Four milliliters per kilogram per hour for the first ten kilos, two for the next ten, one for anything above that. The system automates that math to reduce human error.
Corn
So the "order IV" prompt isn't the system deciding the patient needs fluids. It's the system saying "the protocol says this patient meets criteria for fluids, and here's the calculated rate."
Herman
And then it routes the order. That's where the integration with the EHR comes in. The triage system pulls patient data from the EHR and pushes orders back. The "order IV" message is likely a computerized provider order entry trigger. It either routes to the physician for approval or, depending on the institution's policies, directly to nursing.
Corn
So the tech isn't ordering the IV. The tech is confirming that the protocol's criteria are met, and the system is doing the paperwork.
Herman
The tech is the sensor. The system is the protocol. The physician is the authority.
Corn
That's a clean division of labor.
Herman
It is. And it's been refined over decades. The Schmitt-Thompson protocols are updated annually by a panel of physicians. They're evidence-based, they cover pediatric and adult triage, and they're designed to be conservative. The system would rather over-triage than under-triage. Send someone to the ER who didn't need it, fine. Miss the kid who did, not fine.
Corn
Can you give me a concrete example of what that conservatism looks like in practice? Like, a specific protocol decision where the system deliberately errs on the side of caution?
Herman
Sure. Take the pediatric vomiting protocol. A kid comes in, vomiting, some abdominal pain, but otherwise looks okay. The protocol doesn't just ask "is the kid vomiting." It asks about duration, frequency, whether there's blood, whether there's bile, whether the kid is producing urine, whether the parent reports lethargy. And here's the conservative part: if the answers are ambiguous, if the parent isn't sure whether the kid has been urinating normally, the protocol defaults to the higher-acuity branch. It says, in effect, "treat this as a potential dehydration case until you can rule it out." That's the over-triage bias baked into the tree. The flowchart doesn't let the tech talk themselves out of escalating.
Corn
So ambiguity is treated as risk.
Herman
Ambiguity is treated as risk. That's the design philosophy. And it's not just pediatrics. Adult chest pain protocols do the same thing. A forty-five-year-old with indigestion and a weird feeling in their chest? The protocol doesn't let you call it heartburn. It routes to the cardiac workup until you've got data that says otherwise.
Corn
So if it's mostly rule-based, where does AI actually fit in? And what's coming next?
Herman
AI fits in at the edges. Mednition's KATE is the clearest example. KATE uses natural language processing to analyze nurse notes and predict sepsis risk. It's not replacing the triage decision tree. It's augmenting it. The nurse still triages the patient. KATE reads the notes and says "hey, this pattern looks like early sepsis, consider escalating."
Corn
So it's a second set of eyes, not the first.
Herman
A second set of eyes that never gets tired and has read every sepsis case in the training data. But it's narrow. It's task-specific. It's not reasoning about the patient. It's pattern-matching against a specific outcome.
Corn
And that's the key distinction. The triage logic is rule-based. The AI is a risk score.
Herman
The AI is a risk score. Epic has a Deterioration Index, which is a deployed AI-based risk score. It takes structured EHR data and outputs a number. It's not a diagnosis. It's a probability that this patient is going to deteriorate in the next however many hours.
Corn
So the model isn't saying "this is sepsis." It's saying "this patient looks like the patients who turned out to have sepsis."
Herman
And that's a meaningful difference, because it changes how you use it. You don't treat the score. You treat the patient. The score is a prompt to look closer.
Corn
Which brings us to the regulatory question. Why isn't reasoning-based AI running the triage logic?
Herman
Because clinical decision support is regulated by the FDA as a medical device if it meets certain criteria. Rule-based systems are easier to validate, explain, and audit. You can trace every decision back to a node in the tree. AI models, especially deep learning, are black boxes. You can't explain why the model output that score. You can only say that it did.
Corn
And in medicine, "the model said so" is not a defense.
Herman
It's not. The FDA has cleared some AI-based CDS tools, but adoption is slow. The regulatory framework is still catching up. And the liability question is unresolved. If the system suggests an action and the tech follows it, who's responsible if something goes wrong? The tech? The hospital? The vendor? The physician who signed off on the protocol?
Corn
So the rule-based dominance isn't just technical. It's legal.
Herman
It's legal and it's practical. A rules engine can be audited. You can show a regulator the exact path the system took. With a deep learning model, you can show them the input and the output, and shrug at everything in between.
Corn
What about large language models? Everyone's talking about reasoning-based AI. Could you run triage on an LLM?
Herman
You could, in theory. You could feed the vitals and symptoms into a model and ask it what to do next. But the risk of hallucination is unacceptable for this workload. You need deterministic outputs. You need the same input to produce the same output every time. LLMs are probabilistic. They're designed to be creative. That's the opposite of what you want in a triage protocol.
Corn
So the calculator problem.
Herman
The calculator problem. If the model gets the dose wrong, that's not a benchmark miss. That's patient harm. And the model doesn't know it got it wrong. It just generates the next token.
Corn
Let me push on that a little. Because I've seen demos where people run clinical vignettes through GPT-class models and the outputs look pretty good. They'll correctly identify the Holliday-Segar calculation, they'll flag the red flags. So what's the gap between "looks good in a demo" and "deployable in an urgent care"?
Herman
The gap is the tail. The demo uses clean inputs. A real triage encounter has a parent who says "he's been throwing up since, I don't know, Tuesday? Maybe Monday?" It has a tech who mistypes a weight. It has a kid who's crying so loud the tech can't hear the answer to the question. The model has to be robust to all of that. And more importantly, it has to fail safely. When a rules engine hits an input it doesn't recognize, it defaults to the conservative branch. When an LLM hits an input it doesn't recognize, it confabulates. It fills the gap with something plausible. And plausible is the most dangerous thing in medicine.
Corn
Because plausible sounds right.
Herman
Plausible sounds right. And a tech under pressure, with a waiting room full of people, is not going to second-guess the plausible answer. That's the failure mode. Not the model being obviously wrong. The model being subtly wrong in a way that matches the human's prior.
Corn
So if AI is going to be used, it's going to be narrow. Task-specific.
Herman
Narrow and task-specific. A sepsis prediction model. A deterioration index. A readmission risk model. These are trained on structured EHR data and output a risk score. They don't diagnose. They don't prescribe. They flag.
Corn
And they're already deployed.
Herman
They're already deployed. Quietly. Epic's Deterioration Index is running in a lot of hospitals. Mednition's KATE is running in a lot of emergency departments. The COVID-19 pandemic accelerated adoption of triage protocols and telehealth triage. A lot of systems that were "nice to have" became "must have" when every patient was a potential exposure.
Corn
The quiet deployment is the story. These systems are everywhere, and patients rarely see them.
Herman
Patients see the screen sometimes, like Daniel did. But mostly they see the tech, and the tech sees the screen. The decision support is invisible.
Corn
Which raises the automation bias question. If the system prompts the tech, does the tech stop thinking?
Herman
That's the risk. Automation bias is real. If the system says "check for vomiting," the tech checks for vomiting. But what if the tech would have noticed something else, something the protocol didn't ask about? The protocol is a tool, not a replacement for clinical judgment.
Corn
Is there a documented case of that? Automation bias causing a miss in a triage setting?
Herman
There are studies. The classic finding is in radiology, where CAD systems for mammography caused readers to miss cancers they would have caught on their own, because they anchored on the CAD output. In triage, the concern is similar. If the protocol doesn't ask about, say, a recent tick bite, and the tech is running down the checklist, the tech might not think to ask. The protocol shapes what gets asked. And what doesn't get asked doesn't get documented. And what doesn't get documented doesn't get considered.
Corn
The protocol is a lens. And a lens can also be a blinder.
Herman
A lens can also be a blinder. That's the trade. You gain standardization, you lose some peripheral vision. The mitigation is training. You train the tech to treat the protocol as a floor, not a ceiling. Ask the questions the protocol asks, then ask the question the protocol didn't.
Corn
And alert fatigue.
Herman
Alert fatigue. If the system prompts too often, the tech starts ignoring the prompts. You see this with drug interaction alerts in the EHR. The alert fires so often that everyone clicks through it. The one time it matters, it gets clicked through too.
Corn
The design challenge is not just getting the prompts right. It's getting the frequency right.
Herman
It's getting the signal-to-noise ratio right. A prompt that fires on every patient is noise. A prompt that fires on the right patient is signal. And the difference is the protocol design.
Corn
Which brings us back to the Schmitt-Thompson protocols. They're updated annually. They're evidence-based. They're designed to be conservative. That's the signal.
Herman
That's the signal. And the tech is the human in the loop. The tech confirms the inputs. The tech decides whether the prompt applies. The tech is the one who looks at the patient, not just the screen.
Corn
The system doesn't make autonomous decisions.
Herman
The system prompts and suggests. Humans make the final call. The tech, the nurse, the physician. That's the regulatory framework, and that's the clinical reality. The system is a tool. It's a very sophisticated tool, but it's a tool.
Corn
The sophistication is hidden behind the big buttons.
Herman
The sophistication is hidden behind the big buttons. The UI is simple because the UI has to be simple. The backend is complex because the backend has to be complex. That's the trade.
Corn
Hilbert, you've been quiet. What's your take on all this?

Hilbert: I worked triage in a Level One trauma center. Late seventies, early eighties. We had a laminated card. Red, yellow, green, black. The card had the protocols on it. You'd read the card, you'd tag the patient. That's what Daniel saw. Just on a screen.
Corn
A laminated card.

Hilbert: Still have one. In my wallet. "If patient presents with abdominal pain and vomiting, check for dehydration and consider IV fluids." That's the line. That's exactly what the system told the tech. Just with more buttons.
Herman
The logic is the same. The delivery changed.

Hilbert: The delivery changed. The card didn't talk back. You had to remember to check it. The screen talks back. That's the difference. The screen remembers for you.
Corn
The integration with the EHR. That's new.

Hilbert: That's new. We wrote everything on paper. The chart was paper. The orders were paper. The lab results were paper. The screen pulls it all together. That's the real innovation. Not the buttons. The integration.
Herman
The integration is the hard part. HL7, FHIR, the interfaces between systems.

Hilbert: I don't know about that. I know we had a horse. Triage. Ran at Aqueduct in eighty-four. I put two hundred on him. He came in last.
Corn
You bet on a horse named Triage.

Hilbert: Seemed like a good idea at the time. It wasn't.
Herman
The integration is the innovation. The protocol is the same. The card is the same. The screen just makes it harder to forget.

Hilbert: Easier to track. We never knew if the protocol worked. We just followed it. Now they can look at the data. See who came back. See who got worse. That's the difference.
Corn
The quiet deployment is also a quiet feedback loop.

Hilbert: That's right. The system learns. Not the AI. The institution. They see what works. They update the protocol. The card gets better.
Herman
The card is now a decision tree in software.

Hilbert: The card is a decision tree. The software is a decision tree with a memory.
Corn
The line between rule-based and AI-based is blurring. The rules are getting smarter. The AI is getting more constrained.
Herman
The rules are getting smarter because they're being informed by data. The AI is getting more constrained because it's being regulated. They're meeting in the middle.
Corn
The patient is in the middle.
Herman
The patient is in the middle. The patient sees the tech. The tech sees the screen. The screen sees the protocol. The protocol sees the evidence. And somewhere in there, the decision gets made.
Corn
The question is how much of that decision is human.
Herman
Right now, most of it. The system prompts. The human decides. But the prompts are getting better. And the humans are getting more trusting. That's the trend to watch.
Corn
The trend to watch is trust. How much do we trust the system to tell us what to ask?
Herman
How much do we trust ourselves to ask the question the system didn't.
Corn
That's the open question. Not whether the AI can reason. Whether we can still reason when the AI is reasoning alongside us.
Herman
That's the question. And it's not just triage. It's everywhere decision support is deployed. The system is a tool. The tool shapes the user. The user shapes the outcome.
Corn
Daniel was looking at a laminated card with a memory. And the memory is getting better.
Herman
The memory is getting better. And the card is getting smarter. And the tech is still the one holding the thermometer.
Corn
Thanks to Hilbert Flumingtop, our producer, for keeping the show running.
Herman
If you enjoyed this episode, leave a review on Apple Podcasts or Spotify. It helps other listeners find the show.
Corn
This has been My Weird Prompts. Email us at show at my weird prompts dot com. We'll be back soon.
Herman
See you tomorrow.

This episode was generated with AI assistance. Hosts Herman and Corn are AI personalities.