Audio Coursesby JellypodLesson 07 of 15
Overview
Explore how artificial intelligence is reshaping triage processes in critical care settings. We delve into the strengths, limitations, and real-world impact of AI-driven triage, highlighting safety, flow improvements, and future implications for emergency departments.
Jeremy: Welcome back to the TIME Podcast. Whether you’re listening here at TIME or catching up later, we’re really glad you’ve joined us. This series is produced by Clintix — the team behind Clintix Pro — and the organisers of TIME.
Hamish: TIME exists to give clinicians space to think — particularly about decisions that feel reasonable in isolation, but look very different when you step back and see the system they create.
Jeremy: And today’s topic sits squarely in that space. AI-powered triage is often framed as a way to improve emergency department flow — faster decisions, better prioritisation, fewer bottlenecks.
Hamish: It’s also increasingly being discussed in boardrooms, digital health strategies, and performance conversations — often with the implication that if we just get triage “right”, the rest of the department will follow.
Jeremy: Which is where this gets uncomfortable.
Hamish: Because triage has always been the place where we concentrate risk when the system is under strain. And adding AI into that space doesn’t just change how we assess patients — it changes how responsibility, blame, and expectation are distributed.
Jeremy: So today we’re not asking whether AI can predict risk at triage. We already know it can.
Hamish: We’re asking something harder: when AI-powered triage is introduced into real emergency departments, does it actually improve patient flow — or does it simply make system failure more visible and more defensible?
Jeremy: And that question matters, because if we get this wrong, the cost isn’t theoretical. It shows up as delayed care, moral injury, and risk that’s been identified but not owned.
Hamish: This is not an episode about hype. It’s about judgement — and where judgement still has to sit, even when an algorithm is in the room.
Hamish: Before we even talk about evidence, it’s worth asking why triage keeps attracting AI solutions in the first place.
Jeremy: Because it looks clean. It’s time-bounded, data-rich, and happens before most of the mess downstream becomes visible.
Hamish: Exactly. Triage sits at a liminal point — not quite assessment, not quite treatment — but it carries enormous downstream consequences.
Jeremy: And historically, triage has functioned as a kind of shock absorber for system strain. When beds disappear or staffing thins out, triage absorbs that pressure by stretching waiting times and reprioritising risk.
Hamish: Which is precisely why it’s politically attractive. If you can make triage “smarter”, it feels like you’re acting on crowding without having to touch the harder problems.
Jeremy: But this is the first conceptual mistake. Flow is not a property of triage. It’s an emergent property of the entire system.
Hamish: Triage can decide who waits. It cannot decide where patients go once the system is full.
Jeremy: And in Australia and Aotearoa New Zealand, triage is operating in a context of structural constraint — access block, ambulance offload delays, chronic workforce shortages.
Hamish: So when AI enters triage, it’s not entering a neutral space. It’s entering a pressure valve that already exists to protect the rest of the system.
Jeremy: Which raises an uncomfortable question: are we using AI to improve decision-making — or to make rationing look more objective?
Hamish: That’s not an accusation. It’s a design reality. Any tool that operates at triage will inevitably participate in how scarcity is managed.
Jeremy: And that’s why claims about “improving flow” at triage need to be treated carefully. You can reorder risk without changing throughput.
Hamish: In fact, you can sometimes make the system feel calmer while nothing materially improves for patients.
Jeremy: So from the outset, AI-powered triage shouldn’t be evaluated as a technical upgrade.
Hamish: It should be evaluated as a redistribution of responsibility under constraint.
Hamish: Before we criticise anything, we should be clear about where AI-powered triage actually does perform well — because otherwise this just sounds like resistance dressed up as caution.
Jeremy: Agreed. And the strongest evidence base here is in retrospective prediction. There are now multiple high-quality modelling studies showing that AI can predict certain outcomes at triage better than clinicians alone.
Hamish: A commonly cited example is the 2021 paper by Raita et al., published in Annals of Emergency Medicine. That study used data from multiple emergency departments in the United States.
Jeremy: They trained a machine-learning model using triage-time variables — vitals, demographics, presenting complaint — and asked a very specific question: can we predict hospital admission and critical care requirement earlier than clinicians do?
Hamish: And the answer, statistically, was yes. The model showed strong discrimination, with AUCs that exceeded clinician-only prediction.
Jeremy: Similar findings have been reported in European and UK datasets. When the outcome is binary and system-facing — admission versus discharge — these models are genuinely impressive.
Hamish: But here’s where we need to slow down, because this is where interpretation often goes wrong.
Jeremy: Exactly. These studies tell us what AI can predict, not what it can change.
Hamish: A model that predicts admission accurately does not shorten length of stay unless the system is able — and willing — to act on that prediction.
Jeremy: And in most of these retrospective studies, there’s no intervention. No change to workflow. No capacity unlocked downstream.
Hamish: Which means there’s an implicit leap being made: that better foresight will automatically translate into better flow.
Jeremy: And that leap is not supported by evidence.
Hamish: There’s also a deeper issue here that clinicians tend to pick up on instinctively: these models are trained on historical decisions.
Jeremy: They learn from who we admitted, how we documented, how risk was interpreted in that particular system at that particular time.
Hamish: So when we say “AI outperforms clinicians,” what we often mean is that it reproduces historical clinician behaviour very efficiently — and very consistently.
Jeremy: Which may be valuable, but it’s not neutral. It bakes in local practice patterns, biases, and system constraints.
Hamish: And that matters for equity. If your historical system under-triaged certain populations, the model will learn that.
Jeremy: There’s also a cognitive load issue that rarely gets acknowledged in these papers.
Hamish: Right. Because prediction doesn’t arrive in a vacuum. It arrives as another signal, another alert, another probability score in an already crowded cognitive environment.
Jeremy: So even if the model is “right,” clinicians still have to decide when to trust it, when to override it, and how to integrate it with their own judgement.
Hamish: And under pressure, humans adapt. Sometimes in sensible ways. Sometimes in ways that introduce new risks.
Jeremy: Which brings us to the central tension of this segment.
Hamish: AI triage is very good at telling us what will happen.
Jeremy: But emergency departments run on what we are able to do — not what we can foresee.
Hamish: And confusing those two is where a lot of well-intentioned implementations start to wobble.
Hamish: This is the point where the literature usually thins out — and where real departments start to recognise themselves.
Jeremy: Because up until now, we’ve mostly been talking about what models can do in controlled, retrospective environments.
Hamish: And that’s exactly why the study by Akhlaghi and colleagues at St Vincent’s Hospital Melbourne, published in Emergency Medicine Australasia, is such an important inflection point.
Jeremy: The title alone is worth pausing on: “Evaluation of a machine learning model for predicting hospital admission after deployment in an emergency department.”
Hamish: That word — after — is doing a lot of work.
Jeremy: This wasn’t a development paper. It wasn’t an internal validation. It was a post-deployment evaluation of a live AI triage system embedded into routine clinical practice.
Hamish: Over 77,000 consecutive ED presentations. No cherry-picking. No idealised dataset. Real clinicians, real documentation, real noise.
Jeremy: And importantly, the model didn’t rely on structured data alone. It analysed free-text triage notes, which is what we actually use — and which is notoriously variable.
Hamish: That choice matters, because it exposes the model to the same ambiguity and inconsistency clinicians work with every day.
Jeremy: Performance was clinically meaningful but imperfect — sensitivity and specificity in the low-to-mid 70 percent range, with a strong negative predictive value.
Hamish: Which already tells you this wasn’t a marketing paper.
Jeremy: But the most important finding wasn’t the absolute performance. It was the performance drop compared to the model’s development phase.
Hamish: And that’s where this paper earns its credibility. Because once the model went live, reality asserted itself.
Jeremy: Documentation styles shifted. Case mix evolved. Clinicians adapted their behaviour once decision support was visible.
Hamish: Which is exactly what always happens — but is rarely measured.
Jeremy: And here’s the key thing: the authors didn’t try to explain that away. They didn’t reframe it as “acceptable degradation.”
Hamish: They named it as an expected property of live clinical systems — and argued that ongoing monitoring and recalibration are non-negotiable.
Jeremy: What they also didn’t do is just as important.
Hamish: They didn’t claim reduced emergency department length of stay.
Jeremy: They didn’t claim improved throughput, reduced crowding, or faster ambulance offload.
Hamish: Which, frankly, is refreshing.
Jeremy: Instead, the contribution of this paper is more uncomfortable and more useful: it shows that deploying AI at triage is not an endpoint — it’s the beginning of governance work.
Hamish: And that reframes the whole conversation. Because once a model is live, responsibility doesn’t sit with the algorithm.
Jeremy: It sits with the department, the hospital, and the system that chose to deploy it.
Hamish: Which raises a question we don’t ask often enough: when an AI tool flags risk at triage, who owns the obligation to act?
Jeremy: Is it the triage nurse? The senior clinician? The bed manager? The executive who approved the rollout?
Hamish: The St Vincent’s study doesn’t answer that question — but it forces us to confront it.
Jeremy: And in doing so, it exposes why so many AI triage implementations feel disappointing in practice.
Hamish: Not because the models don’t work — but because the system around them hasn’t decided how responsibility is redistributed once risk is made explicit.
Jeremy: That’s the real shift this paper represents. It moves us from “Can we build this?” to “Are we prepared to own what it shows us?”
Jeremy: At this point, it’s worth being fair. There are studies where AI-supported triage appears to improve flow — and we should name them properly.
Hamish: Yes, but we should also be clear about what kind of flow, where, and at what cost.
Jeremy: The clearest example comes from chest pain. The study most people reference is by Than and colleagues, published in JAMA Internal Medicine.
Hamish: This was conducted across multiple hospitals and evaluated an AI-informed, risk-driven chest pain pathway — not just a prediction model sitting in isolation.
Jeremy: Exactly. The system used AI-derived risk stratification to guide early decision-making within an already established chest pain pathway.
Hamish: And the outcomes were meaningful within that pathway: shorter length of stay for chest pain patients, faster disposition decisions, and fewer unnecessary admissions.
Jeremy: But this is where nuance matters. That success depended on several preconditions.
Hamish: The pathway already existed. It was staffed. There were agreed endpoints. Cardiology, ED, and bed management were already aligned.
Jeremy: AI didn’t create capacity — it aligned patients with capacity that was already there.
Hamish: Which is a very different claim from “AI improves ED flow.”
Jeremy: I agree — but I’d argue that’s still a legitimate form of flow improvement.
Hamish: It is, but it’s a local optimisation. And local optimisation in a constrained system can have unintended consequences elsewhere.
Jeremy: Such as?
Hamish: If you accelerate one cohort without increasing overall capacity, you often displace congestion onto another group — usually lower-acuity or socially complex patients.
Jeremy: That’s fair. You improve median times for one pathway, but the tail gets longer for others.
Hamish: And that’s rarely acknowledged in the headline results.
Jeremy: Another area worth mentioning is consistency rather than speed. The Levin et al. study in Annals of Emergency Medicine looked at AI-supported triage category assignment across multiple EDs.
Hamish: Their primary outcome wasn’t throughput. It was reduction in variability between clinicians and sites.
Jeremy: And they showed that AI support could standardise triage decisions — fewer extreme outliers, more predictable distribution of acuity.
Hamish: Which doesn’t necessarily shorten waiting times.
Jeremy: But it does change the shape of demand.
Hamish: Yes — and that’s where I get uneasy. Predictability is useful, but it can also mask persistent congestion.
Jeremy: Explain that.
Hamish: If your system becomes more predictable but no less congested, leadership can mistake stability for improvement.
Jeremy: So the dashboards look calmer, but the lived experience doesn’t improve.
Hamish: Exactly. That’s a classic moral hazard. We smooth the signal without fixing the problem.
Jeremy: I take that point. But I’d still argue that predictability has operational value — staffing, escalation, senior oversight.
Hamish: I don’t disagree. I just want us to be honest about what kind of value it is — and what it isn’t.
Jeremy: So the synthesis here is that AI-supported triage can improve pathway-level flow and operational predictability.
Hamish: But it doesn’t produce ED-wide throughput gains unless downstream constraints are addressed.
Jeremy: And expecting it to do so sets both clinicians and the technology up to fail.
Jeremy: Let’s pause for a moment, because we’ve covered a lot of ground, and it’s worth being explicit about where the evidence has taken us.
Hamish: Up to this point, nothing we’ve discussed suggests that AI-powered triage is useless.
Jeremy: But equally, nothing suggests it’s a solution to overcrowding.
Hamish: The consistent pattern across studies is this: AI improves how we see risk earlier.
Jeremy: It improves prediction, consistency, and sometimes pathway-level efficiency.
Hamish: But flow — real, department-wide flow — is governed by capacity, staffing, and access block.
Jeremy: And confusing better foresight with better outcomes is where systems get into trouble.
Hamish: Which brings us to the next question: what happens to clinicians when we add more signal to an already overloaded environment?
Jeremy: One of the more compelling arguments for AI at triage isn’t actually about flow — it’s about safety.
Hamish: Specifically, the idea of continuous re-triage. Moving away from triage as a single, static decision.
Jeremy: Several observational studies have looked at AI systems that continuously monitor vitals, labs, and documentation to identify deterioration in patients waiting to be seen.
Hamish: And on paper, that makes a lot of sense. Humans are bad at sustained vigilance — especially in noisy, crowded environments.
Jeremy: Early warning systems have shown benefits in inpatient settings, so it’s not unreasonable to think similar logic could apply in the waiting room.
Hamish: But this is where we need to be careful, because translating that into emergency care is not straightforward.
Jeremy: Why not?
Hamish: Because emergency departments already run at the edge of cognitive saturation. Adding continuous alerts doesn’t automatically improve safety — it can also dilute it.
Jeremy: Right. Alert fatigue isn’t theoretical. It’s something every senior clinician has felt.
Hamish: And clinicians adapt. If every second patient is flagged as “at risk”, the signal loses meaning.
Jeremy: Which raises a design question that rarely appears in AI papers: what happens to human judgement under sustained alert pressure?
Hamish: Exactly. More data doesn’t necessarily mean more clarity. Sometimes it just means more noise.
Jeremy: There’s also a subtle redistribution of responsibility here.
Hamish: Yes — once a system is continuously monitoring patients, expectations shift. If deterioration occurs, the question becomes: why wasn’t it acted on?
Jeremy: Even if no action was possible.
Hamish: Which is dangerous. We risk turning clinicians into the final common pathway for system failure — expected to absorb risk that has been identified but not resolvable.
Jeremy: So while continuous re-triage may improve detection, it can also increase moral and cognitive load.
Hamish: And unless leadership explicitly owns what happens when alerts fire and nothing can be done, that burden falls silently on clinicians.
Jeremy: This is where AI safety narratives can become misleading.
Hamish: Because safety isn’t just about detection. It’s about capacity to respond.
Jeremy: And if response capacity doesn’t exist, detection alone may actually make harm more visible — without making it preventable.
Hamish: This is the point where I think we need to let something go wrong — because in real emergency departments, it does.
Jeremy: Alright. Walk me through it.
Hamish: Picture a large metropolitan ED on a winter evening. The waiting room is full. Ambulance offload delays are already in play. Inpatient beds are tight.
Jeremy: So, a normal night.
Hamish: Exactly. Now add AI-supported triage into that mix. The system is live. It’s been embedded into the electronic medical record.
Jeremy: And it’s doing what it was designed to do.
Hamish: Yes. A middle-aged patient presents with vague but concerning symptoms — abnormal vitals, a concerning triage note. The AI flags them as high risk for admission.
Jeremy: So early risk recognition works.
Hamish: It does. The triage nurse escalates appropriately. The alert is visible. The senior doctor is aware. Everyone knows this patient matters.
Jeremy: But nothing downstream moves.
Hamish: Right. There are no cubicles. No monitored beds. No inpatient beds to pull forward. No staff you can redeploy without creating risk somewhere else.
Jeremy: So the patient waits.
Hamish: And this is the failure mode: the presence of the AI flag creates a false sense of safety.
Jeremy: Because the risk has been identified.
Hamish: Documented. Acknowledged. Visible in the system.
Jeremy: But not resolved.
Hamish: Exactly. Time passes. The waiting room gets louder. Attention is pulled elsewhere.
Jeremy: And eventually, the patient deteriorates.
Hamish: Not because the AI failed. Not because the clinicians ignored the signal.
Jeremy: But because the system had nowhere to put them.
Hamish: Now here’s the uncomfortable part. When this case is reviewed later, the documentation looks good.
Jeremy: The risk was recognised early.
Hamish: Escalation was appropriate.
Jeremy: The AI did its job.
Hamish: Which means the system failure is harder to see — and easier to defend.
Jeremy: That’s the moral hazard.
Hamish: Exactly. AI can turn unresolved risk into well-documented risk, and those are not the same thing.
Jeremy: And this is why ED-wide flow gains remain elusive.
Hamish: Because triage — human or AI-augmented — cannot overcome access block, workforce shortages, or structural congestion.
Jeremy: In fact, safer prioritisation can sometimes concentrate risk, making the consequences of inaction more severe.
Hamish: Which is confronting, but it’s the reality leaders need to design for.
Jeremy: And it reframes the question entirely.
Hamish: Yes. The question isn’t “Did the AI work?”
Jeremy: It’s “Once the AI showed us the risk, who owned what happened next?”
Jeremy: So let’s take this out of the abstract and into reality. You’re back in your department on Monday morning. Someone from executive, digital health, or procurement asks whether AI-powered triage is something your ED should be adopting.
Hamish: And the first question is not “does the model work?”
Jeremy: It’s “what problem are we actually trying to solve?”
Hamish: Because if the dominant problem in your department is access block, inpatient bed scarcity, or workforce exhaustion, AI at triage will not fix your flow metrics.
Jeremy: In fact, it may make those failures more visible without making them more solvable.
Hamish: Which can be politically attractive but clinically dangerous.
Jeremy: I want to push back slightly there. Visibility isn’t inherently bad.
Hamish: I agree — if the system is prepared to act on what it sees. If not, visibility just concentrates moral injury at the front door.
Jeremy: That’s fair. So the second judgement call is where AI triage might genuinely add value.
Hamish: And the evidence is pretty consistent here. AI-supported triage makes most sense in departments that already have functional downstream pathways.
Jeremy: Fast track, short stay, chest pain, ambulatory care, frailty pathways — places where early alignment actually changes what happens next.
Hamish: If no such pathways exist, you’re essentially installing a high-resolution thermometer in a room with no cooling system.
Jeremy: You’ll get more accurate readings — but the fever won’t break.
Hamish: The third issue is how the tool is framed to clinicians.
Jeremy: This matters more than most people realise.
Hamish: If AI triage is sold as an operational solution — “this will reduce crowding” — clinicians will quickly disengage when that promise isn’t met.
Jeremy: Whereas if it’s framed honestly as decision support — something that may improve early risk recognition and consistency — it has a chance of being trusted.
Hamish: Trust is the currency here. Once it’s lost, no amount of recalibration will bring it back.
Jeremy: The fourth judgement call is integration and friction.
Hamish: This is non-negotiable. If the AI tool sits outside the EMR, requires duplicate data entry, or interrupts triage flow, it will fail.
Jeremy: Even if the model is excellent.
Hamish: Especially then — because clinicians will resent it.
Jeremy: The fifth issue, and arguably the most important, is governance.
Hamish: Yes. Once an AI system is live, performance will drift. Case mix will change. Equity effects will emerge over time.
Jeremy: So someone has to own monitoring, recalibration, and review.
Hamish: And we need to be explicit about who that is.
Jeremy: Is it the ED director? The digital health team? The hospital executive who approved the rollout?
Hamish: Because if governance is vague, responsibility defaults downward — to the clinician at the bedside or the triage nurse at the desk.
Jeremy: Which is exactly what we should be trying to avoid.
Hamish: There’s also a liability question here that often goes unspoken.
Jeremy: Say more.
Hamish: If an AI system flags a patient as high risk and no action is taken — not because of negligence, but because of capacity — who carries that risk?
Jeremy: The clinician who couldn’t act?
Hamish: Or the system that created an obligation it couldn’t fulfil?
Jeremy: That question has not been answered in most deployments.
Hamish: And until it is, AI triage risks turning clinicians into shock absorbers for system failure.
Jeremy: So the mature position isn’t enthusiasm or rejection.
Hamish: It’s selectivity. Clarity about purpose. And honesty about limits.
Jeremy: And a willingness to say no — or not yet — if the system can’t support what the tool reveals.
Jeremy: When you step back and look across the studies we’ve discussed — from Raita’s retrospective modelling work in the US, to Than’s pathway-specific chest pain studies, to Akhlaghi’s post-deployment evaluation at St Vincent’s Melbourne — a fairly consistent story emerges.
Hamish: AI-powered triage is not a revolution in emergency care. It’s an incremental evolution.
Jeremy: Its most reliable strengths are not speed or throughput. They’re earlier risk recognition, improved consistency, and better situational awareness under pressure.
Hamish: And where flow improvements do occur, they’re narrow, localised, and dependent on downstream readiness. AI can align patients with capacity — but it cannot create capacity where none exists.
Jeremy: The St Vincent’s experience is particularly instructive, because it forces us to confront something uncomfortable: once AI leaves clean datasets and enters real emergency departments, governance becomes the central problem.
Hamish: Not algorithmic performance. Not accuracy. Governance.
Jeremy: Who owns the signal? Who is responsible for acting on it? And what happens when the system simply cannot respond?
Hamish: That question matters because AI doesn’t just surface risk — it redistributes responsibility.
Jeremy: Once risk is identified earlier, more clearly, and more visibly, the tolerance for inaction changes.
Hamish: And if that redistribution isn’t explicit, it flows downhill — onto the triage nurse, the junior doctor, the consultant on the floor.
Jeremy: Which is how clinicians quietly become shock absorbers for system failure.
Hamish: This is why it’s so important to say this clearly: overcrowding is not a triage problem.
Jeremy: It never was.
Hamish: Triage — whether human, AI-augmented, or hybrid — can prioritise risk, improve safety, and stabilise operations.
Jeremy: But it cannot compensate for access block, workforce shortages, or structural congestion.
Hamish: And when we pretend it can, we’re not innovating — we’re relocating accountability.
Jeremy: So the real question for emergency departments in Australia and Aotearoa New Zealand isn’t whether AI will appear at triage.
Hamish: It already has.
Jeremy: The real question is whether we’re prepared to own what it shows us.
Hamish: Whether we’re willing to match better detection with real authority, real capacity, and real leadership decisions.
Jeremy: Because if we’re not, AI won’t make care safer or faster.
Hamish: It will just make system failure clearer — and easier to defend.
Jeremy: The future of triage may well be AI-augmented.
Hamish: But it will remain, at its core, a profoundly human endeavour — one that still demands judgement, courage, and ownership.
Jeremy: Thanks for staying with us through this episode of the TIME Podcast. We know these aren’t always comfortable conversations, but we think they’re important ones to have.
Hamish: If this episode made you pause, disagree, or rethink something you were previously comfortable with, then it’s done its job.
Jeremy: We’d also like to thank Clintix — not just for creating the TIME Podcast, but for hosting the TIME conference itself and making space for this kind of honest, systems-level discussion.
Hamish: TIME exists because people are willing to engage with complexity rather than look for easy answers, and that’s something worth supporting.
Jeremy: Thanks for listening.
Hamish: We’ll see you in the next episode.
