Artificial Intelligence-Enhanced POCUS in the Emergency Department

A Brief Overview

Artificial intelligence (AI) has hastily transitioned from a theoretical promise of the future to a present-day reality impacting our lives in more ways than we even realize. One area in which the impact is being felt is medicine, a notoriously slow-moving field with regards to adopting new ways of thinking or doing things. Despite this historic resistance to change, the potential upside of this revolutionary technology is unequivocal, leaving even the most skeptical AI critics with just a limited few valid arguments against it. The implementation of AI in medicine thus far has offered improvements in illness diagnosis, provider workflow, and clinical decision making; some of the well-established ways in which AI is utilized include: AI-powered triage systems, medical imaging/lab interpretation, documentation/note writing, and billing and coding. While these examples highlight the more universal uses of AI, there also exist more niche use cases, specifically in the area of another technology that experienced its fair share of skepticism in its early days—Point-of-Care Ultrasound (POCUS).

AI in POCUS: Where We Are Today and the Path Forward

Emergency medicine has historically been at the forefront of adopting point-of-care technologies that enhance clinical efficiency, and over the past several years, POCUS has steadily climbed the hierarchy to position itself as one of the most useful and versatile diagnostic tools, particularly in the emergency department (ED) setting.¹ The ED represents perhaps the most fertile ground for AI-enhanced POCUS implementation due to several key characteristics that distinguish it from other clinical settings. The ED is particularly characterized by its unpredictable patient flow, variation in acuity, and the need for rapid diagnostics and decision-making. This fast-paced, high-pressure environment demands tools that can provide reliable information quickly and accurately, precisely what AI-guided POCUS aims to deliver.

As it stands today, the utility of POCUS is largely influenced by pre-programmed tools intended to assist operators in making sense of obtained images—tools ranging from as basic as a ruler to the broadly used M-mode for dynamic analysis (e.g., lung sliding, fetal heart rate) to more nuanced and exam-specific measurements like the Doppler feature for flow analysis (e.g., ejection fraction, vessel stenosis). Despite these tools, one of the biggest limitations of POCUS is the fact that it is a heavily operator-dependent diagnostic modality. Thus, great efforts have been made to minimize the variability of image quality seen between operators. Many publications have focused on the use of AI in one isolated aspect of POCUS—image interpretation. For example, one study published in late 202⁴ looked at the diagnostic accuracy of cardiac dysfunction by ultrasound fellowship-trained physicians versus an AI algorithm, both tasked with evaluating images obtained by expert sonographers. The results showed similar accuracy in diagnosing cardiac dysfunction between the two groups.² While a major step in the right direction, the majority of similar studies lack the true essence of AI—they are more akin to analytical software masquerading as AI. The other important but largely neglected aspect of POCUS is image acquisition, which requires hours and hours of human experience to become proficient. In contrast, deep learning models can be trained in a fraction of the time using large datasets, e.g., ultrasound images labeled “good” or “poor,” designating the quality of the image. Some current AI models have been trained to identify anatomical landmarks, while others are able to sense the position of the probe relative to the patient’s body. As a real-world validation of this advancement, a study published in early 2025 looked at the proportion of diagnostic-quality images obtained on lung ultrasound by expert sonographers without AI assistance versus non-expert trained health care professionals with AI assistance. The results showed a 98% success rate of obtaining diagnostic-quality images in the non-expert AI-assisted group without a significant difference when compared to the expert sonographer group.³

An early private competitor in this niche field aiming to bring this technology to the mainstream is UltraSight™, which previously completed a prospective multicenter study that compared the quality of echocardiographic exams between expert sonographers and inexperienced nurses and residents (i.e., no prior experience) and the results were unprecedented. With just minimal training, the inexperienced users were able to acquire images of diagnostic quality that approached those of experts.⁴ Currently, UltraSight™, in conjunction with Bristol Myers Squibb, is in the recruitment stage for an upcoming study to expand its research based out of Cleveland Clinic that intends to utilize its Real-Time Guidance software in order to assess its efficacy/reliability in actively guiding non-sonographers in obtaining cardiac stroke volume measurements (i.e., left ventricular outflow tract velocity-time integral [LVOT VTI] measurements) compared to expert sonographers.⁵ As mentioned, different AI models have unique capabilities, like identifying landmarks and sensing probe position. In combining the capabilities of multiple models into one, real-time guidance can be given to the ultrasound operator, such as “tilt the probe 15° right to avoid rib shadowing” or “move the probe 2 cm to the left to visualize the gallbladder” while at the same time auto-adjustments are made to optimize settings such as gain and depth to obtain the most ideal image. To take it a step further, visual representations of probe adjustments that need to be made can be displayed on the very screen being used to obtain the images, which is something that UltraSight™ has already accomplished and has incorporated into its software. The benefits of this type of interactive/generative algorithm are readily apparent for novice operators, but even the most experienced operators can benefit, as other factors like patient habitus and anatomical variations further complicate the ability to obtain adequate imaging.

In light of the promising future that AI-driven POCUS offers, challenges and limitations to its widespread adoption exist, owing to the major concerns regarding AI in medicine overall, which largely center around the ideas of patient privacy, bias (intentionally programmed or not), and effects on the patient-provider relationship.^6,7 Despite these barriers, the trajectory of one small but increasingly important aspect of medicine is clear—we are heading toward a future in which emergency physicians, and ultimately health care providers as a whole, can leverage AI-driven ultrasound to acquire and interpret diagnostic-quality images, regardless of experience or skill level.

References

1. Smallwood N, Dachsel M. Point-of-care ultrasound (POCUS): unnecessary gadgetry or evidence-based medicine? Clin Med (Lond). 2018;18(3):219-224.
2. Gottlieb M, Schraft E, O’Brien J, et al. Diagnostic accuracy of artificial intelligence for identifying systolic and diastolic cardiac dysfunction in the emergency department. Am J Emerg Med. 2024;S0735-6757(24)00537-0.
3. Hswen Y, Abbasi J. AI-guided lung ultrasounds could help nonexpert clinicians acquire "expert-level" images. JAMA. Published online February 21, 2025.
4. Mor-Avi V, Khandheria B, Klempfner R, et al. Real-time artificial intelligence–based guidance of echocardiographic imaging by novices: image quality and suitability for diagnostic interpretation and quantitative analysis. Circ Cardiovasc Imaging. 2023;16(11).
5. Reliability of minimally trained operator's velocity-time integral measurement guided by artificial intelligence VTI (MiniTrainedVTI). ClinicalTrials.gov identifier: NCT06486467.
6. Tyson A, Pasquini G, Spencer A, et al. 60% of Americans would be uncomfortable with provider relying on AI in their own health care. Pew Research Center. Published February 22, 2023.
7. Farhud DD, Zokaei S. Ethical issues of artificial intelligence in medicine and healthcare. Iran J Public Health. 2021;50(11):2169-2176.