AI POCUS in the ED: What It Sees and What It Misses

A second-year resident handed me the probe at 2 a.m. and pointed at the screen. The AI overlay had just labeled an apical four-chamber view as "adequate" and dropped a green check mark next to the automated ejection fraction estimate: 58%. The patient — mid-seventies, hypotensive, on bilevel — looked back at me from the gurney with a respiratory rate of 32. The number was wrong. The patient had a globally hypokinetic ventricle and a pericardial effusion the algorithm had completely failed to flag because the operator had not yet swept through the subxiphoid window. The algorithm could not see what was not yet imaged. The resident was about to chart it as normal.

That moment is the entire conversation about AI-enhanced point-of-care ultrasound in 2026. The tools are now embedded in nearly every cart-based and handheld device on the market. They are useful. They are also, in specific and predictable ways, wrong. The question for emergency physicians is no longer whether to use them, but how to use them without letting one quietly sign off on a missed diagnosis.

After two years of running AI-enhanced POCUS across my own department and reviewing the 2025–2026 literature, I want to walk through three things: what the current evidence actually shows, the four-window failure mode every ED needs to understand, and the clinical workflow I now run on every AI-augmented scan.

What the 2025-2026 Evidence Actually Shows

The published data on AI-enhanced POCUS is finally substantial enough to make real decisions on. A January 2026 narrative review in Diagnostics synthesized the current body of work and was clear about both the promise and the ceiling. AI-assisted lung ultrasound for pneumothorax detection now reports area under the receiver operating characteristic curves above 0.90 against expert interpretation in well-controlled datasets. AI-assisted eFAST detection of free intraperitoneal fluid is approaching expert-level performance. Cardiac AI tools for ejection fraction estimation and pattern recognition can match fellowship-trained reads when the image quality is adequate.

That last clause is the entire ballgame. Image quality is the upstream constraint that the marketing materials almost never lead with. A 2024 prospective ED study published in the American Journal of Emergency Medicine evaluated AI-driven cardiac POCUS for systolic and diastolic dysfunction in patients aged 45 and older. The AI demonstrated 85.7% sensitivity and 94.8% specificity for systolic dysfunction against expert review — strong numbers. Sensitivity dropped to 80% in patients with a body mass index of 30 or greater. That is the single most important number in the paper, and it is rarely cited in vendor materials. The patients in whom POCUS image quality is hardest to obtain are the same patients in whom AI interpretation degrades fastest.

The acquisition-side story is more encouraging. A 2025 study replicated and reinforced earlier JAMA Cardiology data on AI-guided cardiac scanning: with roughly 2.5 hours of training, non-expert users — including nurses and trainees — were able to obtain diagnostic-quality images at a 98% success rate, statistically indistinguishable from expert sonographers. That finding has real implications for staffing, training, and rural access, and it is genuinely the strongest case for AI in POCUS today. AI guides the probe. The clinician still has to know what they are looking for and what is missing.

The Four-Window Failure Mode

Here is the failure mode I see most often, the one that nearly tripped my resident at 2 a.m. AI-enhanced POCUS interprets what is in the frame. It does not interpret what was never imaged. That sounds obvious until you watch a junior operator generate a confident, AI-labeled "normal" cardiac study off two views and chart accordingly.

The four-window cardiac exam — parasternal long, parasternal short, apical four-chamber, and subxiphoid — exists for a reason. Each window catches a different failure pattern. A pericardial effusion may be obvious from the subxiphoid view and silent on the parasternal long. Right ventricular strain shows best in apical four-chamber. Regional wall motion abnormality in the inferior wall may be invisible on parasternal short.

Current AI tools generate confidence scores window by window. They do not flag the absence of windows. If the operator never obtains a subxiphoid view, the algorithm reports on what it has and produces no warning that the exam is incomplete. The chart reads "AI-assisted bedside echo: normal LV function, EF 58%." The clinical reality is an unimaged effusion.

The same pattern holds for lung ultrasound and eFAST. AI lung tools are excellent at calling B-lines and pneumothorax in the zones imaged. They do not detect that you forgot to scan the right upper anterior zone where the small apical pneumothorax actually lives. AI eFAST tools accurately identify free fluid in Morison's pouch when it is in the frame. They do not tell you the pelvic view was skipped.

The fix is not algorithmic. It is operator-side. Treat the AI output as a single data point inside a structured exam protocol that you, not the algorithm, are responsible for completing.

The Clinical Workflow I Run Now

Three rules, applied to every AI-augmented POCUS exam in my department, regardless of operator experience level.

First, the protocol is the protocol. AI does not shorten the required windows. A cardiac POCUS still requires four windows. A lung exam still requires the standardized zones bilaterally. An eFAST still requires all four standard views plus a pelvic. The AI confidence score on any individual window is irrelevant to whether the protocol is complete. Build this into your departmental QA tracker — random sample of AI-assisted POCUS exams every month, audited for protocol completeness, not interpretation accuracy.

Second, the chart documents both the AI output and the operator interpretation as separate elements. "AI-assisted apical four-chamber: EF estimate 58%, confidence high. Operator interpretation: globally reduced systolic function, qualitatively LV ejection fraction 30 to 35%, AI estimate not consistent with operator assessment." That divergence is now captured. If the operator is wrong, the AI estimate is in the chart. If the AI is wrong, the operator interpretation is in the chart. Both protect the patient and both protect the physician.

Third, AI output is never the sole basis for a critical disposition. If your discharge decision rests on the AI saying ejection fraction is preserved, you do not have a discharge decision yet. You have a screening data point that needs an operator interpretation, and in the cases where the disposition matters most — the dyspneic elderly patient, the post-arrest patient, the trauma patient — you need either a confirmatory study by a more experienced operator or a confirmatory imaging modality. The 2026 Diagnostics review I cited above puts this in formal language: AI-enhanced POCUS is decision-support technology, not autonomous diagnostic technology. The ACEP 2024 ultrasound guidelines treat physician interpretation as the documented standard of care, with AI as an adjunct.

Dr. Chet's Take:

In every operational role I have held — running an air medical program, overseeing critical care transport, building telemedicine networks, serving as a State Surgeon — the same command principle has held. You do not delegate the standard of care to a technology. You delegate execution and you maintain oversight. POCUS is no different. The AI overlay on a probe is a force multiplier the same way a flight nurse with a video laryngoscope is a force multiplier. It expands what your team can do at the bedside. It does not replace your responsibility to verify that the exam was complete and the interpretation was correct.

When I trained crews on critical care transport, the unforgivable error was never the difficult case missed under uncertainty. The unforgivable error was the case missed because the protocol was skipped. The same standard applies here. AI-enhanced POCUS does not let you off the hook for a complete cardiac exam. It does not absolve you of subxiphoid views, pelvic views, or standardized lung zones. The algorithm reports on what you imaged. The patient deserves the full study. Train your residents and your APPs that the AI overlay is a quality check on a protocol they still own, end to end.

What You Should Be Doing Now

Audit your last 30 AI-assisted POCUS exams for protocol completeness, not interpretation accuracy. You will find the gap is bigger than you expected. Build a one-line departmental documentation standard separating AI output from operator interpretation, and make it required for every AI-augmented scan. Train every operator — attending, resident, APP — that the AI confidence score reports on the window in front of it and never on the windows that were never obtained. And tell your vendor that quarterly subgroup performance breakdowns by body habitus, age, and sex are a procurement requirement, not a procurement nicety. If they cannot produce them, that is your answer.

AI Won't Wait. Neither Should You.

If you are running AI-enhanced POCUS in your ED today without a written protocol-completeness audit, a separate documentation standard for AI output, and a vendor performance review process, you are exposed. Consider enrolling in my course: AI in Emergency Medicine: Becoming AI Bulletproof. The course walks through the exact POCUS audit template we run, the operator-versus-AI documentation language vetted against current ACEP and AMA guidance, the vendor evaluation checklist for AI-enhanced ultrasound, and the resident training module I use on every new rotation.

Learn more: AI in Emergency Medicine: Becoming AI Bulletproof.

If you're an emergency physician (or any clinician treating patients daily) trying to understand how AI will actually impact your clinical practice — not just the hype — I put together a free practical guide. You can download it here: AI in EM Survival Guide.

Dr. Chester "Chet" Shermer, MD, FACEP is a Professor of Emergency Medicine, TeleHealth, HEMS and Critical Care Transport, and State Surgeon for the Army National Guard. He is the founder of Global MedOps Command and the creator of AI in Emergency Medicine: Becoming AI Bulletproof.

Sources

1. Diagnostics, "AI-Enhanced POCUS in Emergency Care: A Narrative Review," January 2026, https://pmc.ncbi.nlm.nih.gov/articles/PMC12839756/
2. American Journal of Emergency Medicine, "Diagnostic accuracy of artificial intelligence for identifying systolic and diastolic dysfunction in the emergency department," December 2024, https://www.sciencedirect.com/science/article/abs/pii/S0735675724005370
3. EM Resident (EMRA), "Artificial Intelligence-Enhanced POCUS in the Emergency Department," November 2025, https://www.emra.org/emresident/article/Artificial-Intelligence-Enhanced-POCUS-in-the-Emergency-Department
4. American Medical Association, "AMA Principles for Augmented Intelligence (AI) Development, Deployment, and Use," https://www.ama-assn.org/system/files/ama-ai-principles.pdf