Every health profession shows the same gap in training. And the reason it persists is not a failure of teaching.

A patient is quietly deteriorating. The heart rate has climbed twenty beats over four hours. The respiratory rate is up. But the blood pressure has not moved, because the body is compensating, which is what a body does right up until it no longer can.

A nurse, a physician, a respiratory therapist, a paramedic, or a physician assistant is with the patient. The type of professional at the patient’s bedside may change what occurs next, but it does not change what needs to happen first. Somebody has to notice.

One competency, five different names

Nursing calls this clinical judgment. Medicine codifies it as “Entrustable Professional Activity 10: the competency to recognize a patient requiring urgent or emergent care and initiate evaluation and management.” Paramedicine, respiratory therapy, and physician assistant programs have their own language for it. What they are describing, however, is the same thing: noticing that something has changed, understanding what that means, and escalating while there is still time to act.

And one name for what happens when nobody acts

Failure to rescue, first defined by J. H. Silber and colleagues in 1992, is the death of a patient who developed a complication that was, in principle, survivable. It measures not the rate at which complications occur, but the rate at which they are recognized and reversed. 

This distinction matters: initial research found that complication rates were driven largely by patient factors, such as age, comorbidity, or the nature of the surgery, while failure-to-rescue rates were determined by the hospital. Later data confirmed that hospitals in the lowest and highest mortality quintiles had virtually identical complication rates. It turned out that the difference between a safe hospital and a dangerous one is not whether patients deteriorate, it is whether somebody notices in time.

What the evidence shows

Yet future clinicians do not learn this in school. Across two decades, the proportion of newly graduated nurses that met entry-level expectations for clinical judgment went from 35 percent in 2005 to 23 percent in 2015 – and to just 9 percent in 2021. In a 2023 survey of nurse preceptors, novice nurses’ ability to identify which data actually matter scored at the bottom of the eleven categories tested. 

In 2014, the Association of American Medical Colleges (AAMC) named 13 activities every graduate should perform unsupervised on day one of residency, and a ten-school pilot then asked how well curricula taught and assessed each one. Nine were handled well, such as taking a history, building a differential, or writing a note. Four came back absent or underdeveloped in most schools, including EPA 10: “recognize a patient requiring urgent or emergent care and initiate evaluation and management.” Those four share what the other nine do not: every one of them requires a learner to act on a real patient in real time.

Other health professions offer no data at all. A 2024 systematic review of practice readiness found 93 studies, of which nursing accounted for 39. Physician assistant, respiratory therapy, and paramedicine had a single study each. A 2023 scoping review explains why: clinical reasoning is scarcely explicitly taught or assessed across health professions education, resulting in curricula focused on arriving at a correct diagnosis rather than on the reasoning behind it – and leaving reasoning itself as a “black box” for students to explore on their own.

Why nobody teaches this 

The AAMC pilot explains why schools do not teach clinical judgement. Learners rarely encounter actual emergent patients, and when they do, concern for patient safety usually precludes their involvement in the evaluation and management. When a patient is decompensating at three in the morning, nobody hands that assessment to a learner so she could practice. The senior clinician takes over, as they should. Meanwhile, what the learner most needed to practice – noticing the change before it became an emergency – happened hours earlier, on a quiet unit, to a patient nobody had yet flagged.

Clinical judgement is the one competency that cannot be practiced on real patients at scale, because the moment it matters is exactly the moment nobody will let a learner near the patient. That is not going to change, and it should not. Which means the only way to learn this skill before a professional’s first shift on their own is somewhere else, deliberately and at volume. 

Why more cases is not the answer by itself

To be sure, case-based learning works. But consider what it asks of a curriculum. A gunshot wound can deteriorate through hemorrhage, tension pneumothorax, cardiac tamponade, spinal injury, airway compromise, or sepsis three days later – six pathways from a single presentation. How many cases cover this level of complexity? This type of learning works by accumulation, and competency arrives late, if it ever does.

A physiology-based framework functions differently. Oxygen has to get into the lungs, cross into the bloodstream, travel to the tissue, and be used by the cell. Every respiratory presentation a learner will ever meet is a failure at one of these four steps. All the learner needs to ask is which step failed, and then answer that question based on the evidence in front of her. Rather than replace cases, this framework serves as the structure the cases deposit into, turning a collection into a system.\

And the part that has been missing

Still, one problem remains. We want to draw out a learner’s reasoning so we can see the thinking behind the decision and correct it when it is wrong. But reasoning leaves no record. We can see what she did, but we cannot see what she noticed, what she considered and discarded, or what she decided not to say. So, we assess the decision and never see the path that led to it. When she gets it right, we do not know why. When she gets it wrong, we do not know where.

Clinical reasoning is a non-linear, context-driven construct that arguably cannot be assessed at one point in time or with a single method – which means the answer is capturing the reasoning repeatedly, while it is happening. The mechanism is structural rather than technological. At each decision point the learner commits before the data comes back, naming what she thinks is happening and what she expects to find. That commitment is the record that does not otherwise exist: not the answer she landed on but the reasoning that got her there. A wrong conclusion is a dead end, but a wrong hypothesis shows both the faculty and the learner exactly where the reasoning broke down. 

Why I am working with AIBODY

My doctoral research and most of my career since have been about failure to rescue, or why clinicians miss deterioration when the early signs are present albeit subtle. The answer, repeatedly, is that they were never deliberately trained to recognize those signs. 

What drew me to AIBODY is that the problem is identical across every profession they serve, and so is the solution. A physiological framework is not a nursing framework – oxygen crosses an alveolar membrane the same way regardless of which professional is at the bedside. A paramedic, a respiratory therapist, and a physician assistant ask the same question of the same patient: which step has failed, and how do I know?

A system that generates comparable reasoning data across nursing, respiratory therapy, paramedicine, and physician assistant education would not fill a gap in the existing evidence. It would create new evidence where none currently exists.

The patient does not care what is on your badge

A deteriorating patient produces the same signs regardless of which professional walks in. The heart rate climbs. The respiratory rate rises. Something is not right, but the blood pressure is still stable. Whoever is in that room has to notice, work out what it means, and say so while there is still time. That is one competency shared across every profession that touches a patient – and it deserves to be taught deliberately, practiced at volume, and measured precisely for each profession.

References

Amiel, J., Ryan, M. S., Andriole, D. A., & Whelan, A. J. (2022). Core Entrustable Professional Activities for Entering Residency: Summary of the 10-School Pilot, 2014–2021. Association of American Medical Colleges.

Torre, D., Daniel, M., Ratcliffe, T., Durning, S. J., Holmboe, E., & Schuwirth, L. (2024). Programmatic assessment of clinical reasoning: New opportunities to meet an ongoing challenge. Teaching and Learning in Medicine, 37(3), 403–411.

Elvén, M., Welin, E., Wiegleb Edström, D., Petreski, T., Szopa, M., Durning, S. J., & Edelbring, S. (2023). Clinical reasoning curricula in health professions education: A scoping review. Journal of Medical Education and Curricular Development, 10.

Kavanagh, J. M., & Sharpnack, P. A. (2021). Crisis in competency: A defining moment in nursing education. OJIN: The Online Journal of Issues in Nursing, 26(1).

Kennedy, K., Laskowski, J., & Breyman, K. (2023). Clinical judgment in new nurse graduates: Identifying the gaps. International Journal of Nursing Education Scholarship, 20(1).

Silber, J. H., Williams, S. V., Krakauer, H., & Schwartz, J. S. (1992). Hospital and patient characteristics associated with death after surgery: A study of adverse occurrence and failure to rescue. Medical Care, 30(7), 615–629.

Dr. Julie Siemers, DNP, MSN, RN, is the founder of Lifebeat Solutions and creator of the Patient Safety Standard™ curriculum and the Clinical Judgment & Safety Method™. Her doctoral research, grounded in Patricia Benner’s work, focused on failure-to-rescue prevention.

Failure to rescue is a process, not an event – and we teach it backward.

Ask most health professions faculty to teach failure to rescue and they will reach for a case with an ending. A code. A rapid response. A death that could have been prevented. We debrief the outcome, analyze what went wrong, and identify the missed opportunity.

In doing so, we hand learners a timeline that runs backward from the catastrophe rather than forward from the moment the patient’s physiology first began to signal distress.

That is not how clinical reasoning works. Clinical reasoning runs forward, in real time, with incomplete information, under conditions that require the clinician to recognize a trajectory before it has fully declared itself.

The clinicians who fail to rescue are not, in most cases, the ones who ignored an alarm. They are the ones who did not yet understand that the patient was already in a process.

The three phases every clinician needs to see

The physiological arc of deterioration follows a recognizable pattern. I call it the compensation-decompensation-irreversible failure pathway, and it sits at the center of the Patient Safety Standard™ curriculum for one reason: it is the most clinically actionable thing a learner can carry onto a unit.

In the compensation phase, the body is working. The respiratory rate climbs quietly. The heart rate edges up. Blood pressure holds, and may look entirely normal, because compensatory mechanisms are still intact. A patient in this phase does not look sick. They may be anxious, mildly confused, or simply “not quite right.” The aide mentions it. The therapist mentions it. The family mentions it. Someone documents a set of vital signs that, taken individually, cross no threshold at all.

This is the window. This is where rescue happens, or fails to.

In the decompensation phase, the ability to compensate begins to fail. Subtle trends become unmistakable. Mental status changes. Values begin to cross thresholds. This is the phase where most clinicians recognize that something is wrong - and it is also the phase where the window for easy intervention is already closing. What feels like early action here is late action relative to where the process began.

In the irreversible failure phase, the patient is in crisis. The rapid response has been called. The rescue, if it comes, will be harder, costlier, and less likely to return the patient to baseline. Some patients will not be rescued at all.

We spend most of our clinical teaching time in the third phase. That is not where the learning that saves lives needs to happen.

What faculty miss, and why

Across seventeen years in health professions education, here is what I have observed: faculty are exceptionally skilled at teaching what went wrong. Root cause analysis, case debrief, post-event review – we are trained for this, and we do it well. What we do less well is teaching learners to read a patient who has not yet gone wrong in any documentable sense.

The reason is structural. Our cases tend to begin at the point of deterioration: the patient presents with these findings, the clinician does these things, the outcome is this. The compensation phase – the quiet period of escalating physiological distress before the numbers tell the story – is either omitted entirely or treated as backstory.

When we do that, we inadvertently teach that recognition begins when the data becomes obvious. But the clinician who rescues a patient does not wait for obvious. She recognizes a trajectory. She sees that a respiratory rate of 16 yesterday is 20 today, and asks what that trend means before it becomes 26 at two in the morning. That is a teachable skill, but it is not being systematically taught, in any profession.

How to reframe every case you already use

The fix is not a new curriculum, but a reframe that can be applied to every case already in your rotation.

Start in the compensation phase. Before you name the diagnosis, before you reveal the deteriorating values, put learners in the room with a patient who looks nearly fine. Give them the subtle cues: the slightly elevated heart rate, the mild restlessness, the family’s unease. Then ask: what is this patient’s biggest risk right now, and what is the earliest sign that risk is becoming real?

Let them sit with uncertainty. Let them reason through incomplete data. The cognitive discomfort of that exercise is precisely the mechanism that builds pattern recognition.

Then advance the timeline. Show them what compensation looks like as it becomes decompensation. Ask: at what point did the window open? When did it close? What would someone have needed to recognize in order to act inside it?

Finally, debrief the process rather than the outcome. Shift the question from “what went wrong” to “where was the earliest decision point?” That is the question that changes how a clinician thinks at the bedside, and it is the question that almost never gets asked in a traditional debrief.

Why this belongs to every profession

None of this is specific to any one discipline. A patient in compensated shock produces the same findings regardless of which health professional walks into the room. Oxygen crosses an alveolar membrane the same way whether the person at the bedside is a nurse, a respiratory therapist, a paramedic, a physician assistant, or a physician in training.

What differs is the vocabulary each profession uses for the same competency and, too often, whether anyone teaches it at all. A 2023 scoping review across health professions education found that clinical reasoning is scarcely explicitly taught or assessed, leaving curricula focused on arriving at correct diagnoses rather than on training the reasoning process itself.

The compensation window is where that reasoning either happens or does not. It is also the part of the timeline that is hardest to practice on real patients, because a learner is rarely present for the quiet hours before anything has declared itself. And when deterioration does declare itself, patient safety quite properly keeps learners out of the management.

Which is why this competency has to be built deliberately, in simulation, at volume. It is not going to be picked up on a unit by chance.

What practice-ready actually means

A practice-ready clinician does not merely know the signs of sepsis. She has practiced, repeatedly, the cognitive act of recognizing a sepsis trajectory from data that does not yet look like sepsis. She has done that work in cases, in simulation, in debrief, under conditions ambiguous enough to require genuine thinking.

That kind of preparation is not built by covering more content. It is built by asking the right question at the right point in the clinical story: what is happening to this patient’s physiology right now, and where is it going?

The compensation-decompensation-irreversible failure pathway gives every clinician a map. The job of faculty is to teach learners to read that map before the patient reaches the third phase.

That is the shift. And it does not start with a revised curriculum. It starts with how we frame the first five minutes of every case we already use.

Dr. Julie Siemers, DNP, MSN, RN, is the founder of Lifebeat Solutions and creator of the Patient Safety Standard™ curriculum and the Clinical Judgment & Safety Method™. Her doctoral research focused on failure-to-rescue prevention. She serves on the AIBODY Scientific Advisory Board.

AI can now build an anatomy viewer in a weekend. So is healthcare education about to be flooded with cheap AI slop, or are software providers and publishers facing a genuine day of reckoning? 

Vibe coding will transform healthcare education, but easier software development does not make expertise, validation, or trust optional. AI-assisted development has advanced rapidly. Describe an idea to a coding agent, and within hours you may have a product-like prototype. Healthcare education is a compelling use case. 

A recent viral example used GPT-6 Astra to create an interactive 3D anatomy viewer featuring over 2,000 structures. It showed how AI can rapidly prototype work that once required substantial front-end and 3D expertise. AI did not generate the anatomy. The app draws on BodyParts3D, an established anatomical dataset. AI sped up the software, not the science beneath it. That distinction matters. 

Vibe coding is most powerful in the hands of educators 

I am optimistic about the use of vibe coding in healthcare education. Experienced educators can now turn ideas into working tools without waiting months for a vendor's roadmap. They are creating anatomy apps, ECG tools, simulations, and tailored learning experiences that reflect their teaching approach.  

This is one of generative AI’s most promising possibilities: helping subject-matter experts translate their knowledge into software much faster. Sometimes, a focused tool is all that’s needed, whether it’s an experiment that explains a difficult mechanism, a scenario that addresses a recurring misconception, or a visualization that no commercial platform offers. It does not need to reach 100,000 learners or have an enterprise sales team. It just needs to make tomorrow's class better. That is valuable work. 

But a classroom tool and a platform solve different problems 

The mistake is assuming that because we can now prototype educational experiences quickly, we have solved everything needed to deploy them at scale. Educators building tools for their own classes can make sensible trade-offs. They know their learners, can explain problems, and can modify a tool overnight. They may not need complex administration, analytics, or permissions. 

Products deployed across institutions are different. They raise questions about security, privacy, accessibility, identity, testing, monitoring, support, and governance. Formal education may also require LMS integration, LTI, SSO, roles and permissions, assignment management, grade pass-back, and institutional security reviews. 

These requirements are not glamorous, and AI will make many of them easier. We should not assume today's technical barriers will last forever. Managed infrastructure has already reduced complexity, and AI will increasingly automate testing, integrations, deployment, and security. Small teams will likely build products at a scale that once required larger organizations. 

The argument is not simply that production software is hard. The more interesting question is: “What still needs to be trustworthy when producing software becomes easy?” 

A 3D anatomy viewer is not automatically an anatomy education product 

Stay with anatomy for a moment. 

Interactive 3D anatomy is not new. Google introduced what eventually became Zygote Body more than fifteen years ago, and today there are numerous free and commercial ways to explore a digital human body. 

That does not make a new anatomy viewer pointless. AI enables much more specific experiences for a particular cohort, interaction, language, curriculum, or teaching method. Still, the ability to rotate a beautiful digital body is not a new educational proposition. 

Once we move from exploration to teaching, much harder questions arise. 

What exactly should this learner know? Which structures are relevant at this level, and what should they be asked to do with them? How can anatomy be connected to physiology? How can we tell whether learning has actually taken place? What happens when the learner gets something wrong? How does the activity fit into the broader curriculum? Who validates the material, and who is accountable if something is incorrect? 

Those are not primarily software questions. They are questions of curriculum, assessment, instructional design, and domain expertise. A 3D model can be a powerful educational surface. It is not automatically an educational product. 

Healthcare content can look convincing and still be wrong 

This matters especially as generative AI creates anatomical illustrations, 3D assets, clinical imagery, explanations, and simulations. Healthcare education demands exceptional accuracy. An image can look impressive and still be anatomically wrong. A simulation can feel believable and still be mechanistically incorrect. An explanation can sound authoritative and still be subtly flawed. 

Novice learners are the least equipped to spot these mistakes. Still, textbooks, publishers, educators, and software vendors make mistakes too. AI may also improve validation. Automated checks, structured terminology, consistency tests, source-linked claims, and continuous review could make validation more rigorous than traditional publishing ever was. 

Validation is essential. The key distinction is not between human- and AI-generated content but between validated and unvalidated content. Generative AI is a powerful production tool, not an anatomical authority, physiologist, or editorial board. It can create a beautiful heart, but someone must verify it is the right one and decide what learners should do with the information. 

Maybe we really will “one-shot” some of healthcare education 

There is an even more disruptive possibility: educational tools may not need to become permanent software products. 

Imagine a student struggling with renal physiology asks AI to create an interactive visualization showing how changes in afferent and efferent arteriolar resistance affect the glomerular filtration rate (GFR). The AI creates it. The student spends twenty minutes with it, grasps the concept, and moves on. The software may never be used again. 

Educators could generate activities for a single lecture, and students could create personalized study tools. Institutions might provide a secure foundation while learning experiences are generated on demand. That shift will likely turn entire categories of educational software into commodities. We should not resist that shift. We should build for it. 

The opportunity is bigger than one-shotting software 

The future will not be defined solely by educators building their own tools or traditional vendors making products as they always have. It will include both, and everything in between. Anatomists will prototype new ways to teach structures, physiologists will turn mechanisms into interactive experiments, and nurse educators will develop scenarios drawn from learners’ recurring challenges. 

Clinicians will create cases from practice, students will build tools for their learning needs, and product teams will use the same technologies to develop sophisticated platforms faster. Technology keeps getting cheaper, so the scarce resources will be domain expertise, instructional design, validated content, provenance, and accountability. 

Vibe coding has lowered the barrier to creating software, and AI may eventually make nearly every technical component easier to produce. But healthcare education standards should not fall. When creation becomes effortless, knowing what to trust matters more than ever. 

Part 2: Piracy or Abundance? 

That raises a tougher question: if we use AI to study, transform, or create educational tools from a publisher’s textbook, does buying access alone entitle us to do so? And if the models themselves were trained on proprietary material, are we nearing AI's own “Napster moment,” when technology outpaces the rules for ownership, reuse, and compensation?