AI Pulmonary Embolism Alerts on CTPA: What Gets Faster, What Doesn't

At 0215 your phone shows a triage notice: suspected pulmonary embolism on the CT pulmonary angiogram you ordered 20 minutes ago. The patient is a 61-year-old retired welder with pleuritic chest pain, a heart rate of 112 and an oxygen saturation of 93% on room air. His blood pressure holds at 128/80. The notice carries a small compressed preview image, and the radiologist has not opened the study yet.

That notice moved his scan to the top of the radiologist's list, and so far that is all it has done. In the studies we have, these tools shorten the wait in the reading room, and none of the studies that measured it found faster anticoagulation.

What the PE flag is cleared to do

Many PE alert tools for CT pulmonary angiography (CTPA) are cleared in the United States as radiological computer-aided triage and notification software, a Class II device type defined in 21 CFR 892.2080 [1]. Such a device helps prioritize images and notifies a designated list of clinicians that a time-sensitive study is waiting. It does not remove cases from a reading queue, and it operates in parallel with the standard of care, which remains the default for every case [1].

Two things follow for the emergency physician. A flag means the software saw something worth a faster look. No flag means the scan waits its normal turn and still gets a full read.

How accurate the algorithms are

The best pooled estimate comes from a 2025 meta-analysis of 28 studies, most of them on CTPA [2]. In external validation, the test that matters for a hospital buying a tool someone else trained, pooled sensitivity was 0.89 and specificity was 0.88 [2]. Both figures dropped from internal validation, and heterogeneity between studies was high, so treat them as a range rather than a property of the tool in your building [2]. About 47% of the external-validation scans showed PE, several times the rate on an ED CTPA list, so these figures cannot tell you how many flags will be false.

A French emergency teleradiology group compared one commercial algorithm with emergency radiologists in 1,202 patients with suspected PE and a prevalence of 15.8% [3]. The algorithm was more sensitive (92.6% against 90%) and found 19 emboli that the radiologists missed. The radiologists were more specific (99.1% against 95.8%), and their positive predictive value was 95% against the algorithm's 80.4% [3]. One flag in five was a false alarm. The radiologists' advantage in accuracy disappeared on scans with poor-to-average contrast injection, where the algorithm helped most [3].

At University Hospital Basel, the algorithm's sensitivity in routine use was 79.6% and its specificity 95.0% [4].

What gets faster and what does not

The clearest gain is the time a positive scan waits for a radiologist. In a prospective study at the University of Alabama at Birmingham, mean wait time for PE-positive studies fell from 21.5 to 11.3 minutes with AI triage [5]. Radiologist accuracy (97.6% without AI, 98.6% with it) and miss rate (12.3% without, 6.1% with) did not differ significantly, and examination and report turnaround times did not improve [5].

At UT Southwestern, worklist reprioritization cut mean report turnaround for PE-positive CTPAs from 59.9 to 47.6 minutes, almost all of it wait time [6]. Broken down by priority during regular hours, wait time for routine-priority orders fell from 43.7 to 15.3 minutes, while stat and urgent orders showed no significant change [6].

Downstream of the reading room, the evidence is thinner and observational. Basel measured the emergency department side directly, comparing periods before and after an electronic notification system and then nine months after the algorithm went live. There was no significant reduction in report reading time, report communication time, time to anticoagulation or patient turnaround in the ED [4].

Two 2026 before-and-after cohorts followed patients past the report. At a Cincinnati hospital, among 124 patients who went on to catheter-directed therapy, median time from diagnosis to intervention fell from 1,360 to 1,224 minutes after an AI platform joined the PE response team, while time to anticoagulation did not change significantly (76 against 61.5 minutes) [9]. Time to diagnosis was measured to the AI alert afterward but to the final report before, which favors the AI period, and one author consults for the vendor [9]. At a German university hospital, among 424 included patients with symptomatic central or segmental PE, response-team activation rose from 9.2% to 84.5%, time from CT to diagnosis fell from 2.1 to 1.3 hours, and time from diagnosis to intervention fell by half; hospital and ICU length of stay and 30-day mortality did not change significantly [10]. In both, the algorithm arrived with a new response-team workflow, so the results describe the whole pathway. A faster report helps only if someone sees it and acts on it.

Incidental PE on scans ordered for something else

The strongest result in this field comes from scans nobody ordered to look for PE. At the Netherlands Cancer Institute, AI screened 11,736 routine contrast chest CTs in oncology patients for incidental emboli [7]. With AI worklist prioritization, the radiologists' miss rate fell from 44.8% to 2.6%, and median detection-and-notification time fell to 87 minutes, against 7,714 minutes in routine workflow [7].

That setting was an outpatient oncology backlog with an incidental PE prevalence near 1% [7], which is far from an ED CTPA list. When the question on the requisition is something other than PE, the readers in that study missed close to half of the emboli; with the algorithm flagging those scans, they missed one in 38. Ask radiology whether your tool screens every contrast chest CT, because that decides who gets a call about an embolus nobody suspected.

A found clot still needs a treatment decision, and in patients with cancer the default is treatment. ASCO's 2023 guideline says incidental PE should be treated in the same manner as symptomatic VTE, and that an isolated incidental subsegmental PE is a case-by-case decision weighing the benefit and bleeding risk of anticoagulation [12]. The 2026 multisociety acute PE guideline gives subsegmental PE its own category because triage and the decision to anticoagulate can differ with clot location [11]. When the algorithm's extra find is a single small clot, the software has handed someone a judgment call.

What to do in your department

Find out who receives the notification. The regulation says only that the device notifies a designated list of clinicians [1]. In some hospitals only radiologists see the flag; in others, the ordering physician gets the preview on a phone. Write down who gets the notice, who owns the callback, and how fast. If emergency physicians are on the list, they also need a written rule for what the notice means.

Treat on the read. A flag carries roughly a one-in-five false-positive rate in published real-world use [3]. Your decision to anticoagulate before imaging confirms the diagnosis rests on the same pretest probability and bleeding risk it rested on before the software arrived. A patient who becomes hypotensive needs bedside echocardiography and your hospital's PE response pathway now, whatever the worklist says.

Ask for local monitoring. Mayo Clinic tested a day-by-day monitor that compares the algorithm's output with a language-model reading of the final radiology reports across 11,999 CTPAs; 3.1% of studies disagreed [8]. In simulations, a 2% to 3% drop in specificity produced a two- to threefold rise in disagreements [8]. Our pieces on AI change control, who owns the alert and AI brain-bleed alerts cover the same ground for other tools.

Key Takeaways

• PE triage software is cleared to reorder the radiologist's worklist and send notices; every scan still gets a full read, and the report decides treatment.
• In a 1,202-patient real-world cohort, one in five AI flags was a false positive (positive predictive value 80.4%), while the algorithm also found 19 emboli the radiologists missed.
• AI triage cut the reading-room wait for PE-positive scans by about 10 to 12 minutes in two US studies; in one, the significant gain during regular hours was in routine-priority orders, not stat or urgent ones.
• Where time to anticoagulation was measured, it did not change; two 2026 observational cohorts tied AI-supported response-team workflows to faster diagnosis and intervention, with no significant change in length of stay or 30-day mortality where those were measured.

FAQ

Can I start anticoagulation based on an AI pulmonary embolism flag?

The flag alone is not a diagnosis. The software is cleared for triage and notification and runs in parallel with the radiologist's read [1], and in one real-world cohort about one flag in five was a false positive [3]. Use the notice to ask for a fast read. Whether to anticoagulate before confirmation is the same decision it was before the software existed: pretest probability, bleeding risk and how sick the patient is.

How accurate is AI at detecting pulmonary embolism on CTPA?

A 2025 meta-analysis, mostly of CTPA studies, pooled external-validation sensitivity of 0.89 and specificity of 0.88, with wide variation between studies [2]. About half of those validation scans were positive, so the figures do not predict your false-flag rate. In real-world use, sensitivity was 79.6% in Basel [4] and 92.6% in a French multicenter cohort [3].

Does AI triage of CT pulmonary angiograms shorten ED length of stay?

It has not been shown to. In two US studies, it shortened the wait before a radiologist opened PE-positive scans by about 10 to 12 minutes [5,6]. In Basel, ED turnaround and time to anticoagulation did not change [4]. A 2026 German cohort that paired AI with a response-team workflow found faster diagnosis and intervention, but hospital and ICU length of stay did not change significantly [10].

Dr. Chet's Take

I have practiced emergency medicine for more than 25 years at a Level I academic trauma center, and this article gets the sequence right. The software moves a scan up a list. It does not read the scan, and it does not move the patient. The temptation at 2 a.m. is to treat the phone notice as a result, because it arrives first and it sounds certain. As a HEMS medical director, I also think about the patient whose scan is read at one hospital and whose catheter lab is at another, and a notice is only as fast as the person who receives it and knows what to do with it. The studies say the reading-room savings are real and modest, and the downstream gains appear only where a hospital rebuilt its response team around the alert. The argument holds: buy it for the reading room, and expect the reading room to change.

That being said, the article understates one thing. The ten minutes these tools save are easy to lose on our side of the wall. A report that posts to the chart and waits forty minutes for someone to open it has saved nobody anything. The honest answer is that most departments have never measured the gap between report posted and physician aware, and I suspect it is often longer than the gap the software closed. A second point the evidence does not touch: none of this changes who should get the scan. Pretest probability and the decision rules decide that. An algorithm on the back end does nothing for a loose ordering pattern on the front end, and a low-yield CTPA read faster is still a low-yield CTPA.

If you are leading an emergency department, ask radiology two questions this month. Who receives the PE notice, and what is our local positive predictive value for it? Then measure your own interval from posted report to first anticoagulant dose, because that is the clock your patients feel, and it belongs to you, not to the vendor. If your physicians are on the notification list, agree in writing on what a notice on a stable patient triggers: a call to the reader, a look at the patient, and nothing automatic. If you are the attending, use the notice to pick up the phone and call the radiologist. The report is the result, and the patient in front of you is still your decision.

— Chester Shermer, MD, FACEP | Emergency Medicine, 25+ Years Clinical Experience | State Surgeon

Continue Your Training

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 the AI in EM Survival Guide here.

Browse All Courses at Global MedOps Command

Relevant Reading on Global MedOps Command:

How to Avoid Becoming an AI Casualty — Dr. Shermer's guide to navigating AI tools in clinical and operational settings without compromising judgment or patient outcomes.

Emergency Department Efficiency Playbook — Practical systems for throughput, triage optimization, and operational efficiency.

Read more from Dr. Shermer on Medium

Connect with Dr. Shermer: LinkedIn — Chester "Chet" Shermer, MD, FACEP.

References

1. US Code of Federal Regulations. 21 CFR 892.2080, Radiological computer aided triage and notification software. Cornell Legal Information Institute
2. Farzaei A, Hajzeinolabedini F, Sharif Kashani B, et al. The Role of Artificial Intelligence in Diagnosing Pulmonary Embolism: A Systematic Review and Meta-analysis. Arch Acad Emerg Med. 2025;13(1):e86. PubMed 41664805
3. Cheikh AB, Gorincour G, Nivet H, et al. How artificial intelligence improves radiological interpretation in suspected pulmonary embolism. Eur Radiol. 2022;32(9):5831-5842. PubMed 35316363
4. Schmuelling L, Franzeck FC, Nickel CH, et al. Deep learning-based automated detection of pulmonary embolism on CT pulmonary angiograms: No significant effects on report communication times and patient turnaround in the emergency department nine months after technical implementation. Eur J Radiol. 2021;141:109816. PubMed 34157638
5. Rothenberg SA, Savage CH, Abou Elkassem A, et al. Prospective Evaluation of AI Triage of Pulmonary Emboli on CT Pulmonary Angiograms. Radiology. 2023;309(1):e230702. PubMed 37787676
6. Batra K, Xi Y, Bhagwat S, et al. Radiologist Worklist Reprioritization Using Artificial Intelligence: Impact on Report Turnaround Times for CTPA Examinations Positive for Acute Pulmonary Embolism. AJR Am J Roentgenol. 2023;221(3):324-333. PubMed 37095668
7. Topff L, Ranschaert ER, Bartels-Rutten A, et al. Artificial Intelligence Tool for Detection and Worklist Prioritization Reduces Time to Diagnosis of Incidental Pulmonary Embolism at CT. Radiol Cardiothorac Imaging. 2023;5(2):e220163. PubMed 37124638
8. Sorin V, Korfiatis P, Bratt AK, et al. Using a Large Language Model for Postdeployment Monitoring of FDA-Approved Artificial Intelligence: Pulmonary Embolism Detection Use Case. J Am Coll Radiol. 2025;22(11):1404-1414. PubMed 40602699
9. Austin E, Reichard A, Muck P, et al. Adding artificial intelligence to the pulmonary embolism response team improves time to diagnosis and treatment of pulmonary embolism. J Vasc Surg Venous Lymphat Disord. 2026;14(6):102598. PubMed 42570761
10. Messiha D, Nensa F, Haubold J, et al. Real-world implementation of AI-assisted PERT-workflow and its association with care processes and budget impact: a retrospective cohort study. Eur Heart J Digit Health. 2026;7(9):ztag165. PubMed 42936677
11. Creager MA, Barnes GD, Giri J, et al. 2026 AHA/ACC/ACCP/ACEP/CHEST/SCAI/SHM/SIR/SVM/SVN Guideline for the Evaluation and Management of Acute Pulmonary Embolism in Adults. Circulation. 2026;153(12):e977-e1051. PubMed 41712677
12. Key NS, Khorana AA, Kuderer NM, et al. Venous Thromboembolism Prophylaxis and Treatment in Patients With Cancer: ASCO Guideline Update. J Clin Oncol. 2023;41(16):3063-3071. PubMed 37075273

Related simulation training: EM-Sim · Books by Dr. Shermer