Preserving Clinical Skills in the Era of AI-Enabled POCUS

Alphonce Oyugi, MD1*; Titus Kahiga, PhD2; Kotieno Fredrick, MD2

(1) Global Health Institute, Duke University, Durham, NC, USA

(2) School of Health Sciences, Kenyatta University, Nairobi, Kenya

*Corresponding Author: Dr. Alphonce Oyugi (email: alphonce.oyugi@duke.edu)


Download article PDF – POCUS Journal 2026;11(2):17-20

DOI: https://doi.org/10.24908/pocusj.v11i02.21321


Abstract

Artificial intelligence (AI) is rapidly transforming point-of-care ultrasound (POCUS), with applications spanning image acquisition guidance, automated interpretation, and workflow optimization. While these advances promise to democratize bedside imaging and reduce operator dependency, they simultaneously raise urgent concerns about deskilling, misskilling and never-skilling. This opinion article examines how AI-enabled POCUS may alter the clinical skill landscape, outlines the distinct risks of deskilling, misskilling, and never-skilling in ultrasound practice, and proposes a framework of deliberate strategies to safeguard clinical competency. We argue that the choices made now about how AI is integrated into POCUS training and practice will determine whether these tools elevate bedside diagnostic capability or quietly erode the manual and cognitive skills that define it.

Introduction

Point of care ultrasound (POCUS) has transformed bedside medicine, enabling clinicians across emergency medicine, critical care, internal medicine, and primary care to perform rapid, real-time diagnostic assessments. Its power has always resided in the hand-eye-brain loop of the operator: the tactile skill of probe manipulation, the trained eye for image optimization, and the clinical reasoning that is required to integrate POCUS findings with the patient’s presentation. These competencies are hard-won, requiring structured training and sustained deliberate practice.

Artificial intelligence (AI) is now entering POCUS use at a higher speed than ever imagined. AI applications in POCUS currently span three domains: acquisition assistance (guiding probe positioning and view quality), automated interpretation (image analysis for pathology detection), and workflow optimization (integrating findings into clinical decision pathways) [1]. The AI medical imaging market is valued at approximately $1.15 billion in 2025 and projected to reach $2.57 billion by 2034, driven by rapid adoption of AI-enabled ultrasound systems across clinical settings [2]. U.S. Food and Drug Administration (FDA) submissions for AI medical devices more than doubled between 2020 and 2024, from 114 to 235 submissions. Manufacturers have also now embedded deep learning directly into handheld POCUS devices, placing algorithmic assistance at the first moment of clinical contact [3].

Yet as AI assumes a growing role in POCUS, a foundational question emerges: what happens to the clinician’s own skills? The concern is not hypothetical. On June 1, 2009, Air France Flight 447 crashed into the Atlantic Ocean. All 228 people aboard were killed. The official inquiry highlighted a critical dimension of the automation paradox: when the autopilot was disconnected, pilots with minimal recent manual flying time logged were unable to respond effectively [4]. Heavy reliance on automation had diminished their capacity to intervene when the system failed. Health care is approaching a parallel inflection point.

The literature on colonoscopies offers the earliest clinical signal. A multicenter observational study demonstrated that gastroenterologists and surgeons who used an AI polyp detection system for three months showed measurable declines in adenoma detection rates when AI support was subsequently withdrawn. Their post-AI performance fell below pre-AI baseline, with an estimated 21% decrease in adenoma detection rates, potentially translating to approximately 18% more missed cancer cases [5, 6]. Analogous risks are plausible and arguably amplified in POCUS, where the skill set spans manual dexterity, spatial reasoning, real-time pattern recognition, and integrated clinical judgement. This editorial examines how AI-enabled POCUS may alter the clinical skill landscape, outlines the forms that skill attrition may take, and proposes strategies to ensure AI augments—rather than erodes—the POCUS competencies that define safe bedside practice.

The Deskilling Effect

Concerns about AI-related deskilling are no longer theoretical. As Topaz and colleagues articulated in a commentary in Annals of Internal Medicine, there are hundreds of clinical trials that examine whether AI improves detection during use, yet relatively few investigate what happens to clinician skills when AI is removed or unavailable [4]. The mechanism appears to be cognitive offloading: eye-tracking studies show that clinicians using AI exhibit reduced visual scanning. They stop active searching of the visual field and instead wait for the algorithm to highlight abnormalities. When AI is removed, clinicians accustomed to this more passive verification struggle to reengage active searching [4].

This phenomenon extends across specialties. In mammography, radiologists at all experience levels followed incorrect AI suggestions, demonstrating that expertise provides no immunity to automation bias [6]. In 2025, a randomized trial of diagnostic reasoning, clinicians using a large language model deliberately seeded with errors showed a 14-percentage point reduction in diagnostic accuracy compared with those using accurate AI [7]. A mixed-methods review of AI-induced deskilling in medicine identified four distinct types of deskilling. These included technical skill erosion, decision-making heuristic atrophy, moral responsibility diffusion, and interpretive capacity loss. The review further documented a consistent causal chain from automation bias to complacency to disengagement to skill erosion [8].

The risk of deskilling intersects with several distinctive features of the POCUS skill set. Unlike static image interpretation, POCUS is a real-time, operator-dependent procedure in which image quality is itself a product of skill. AI probe placement assistance may prevent novice users from developing the proprioceptive intuition for acoustic windows. Automated view labelling may short-circuit the spatial reasoning that underpins confident, independent scanning. As AI-enabled POCUS devices proliferate, clinicians risk accumulating scan volume without accumulating the perceptual depth that volume is intended to generate.

Three Trajectories of Skill Loss

Three distinct trajectories of skill loss—deskilling, misskilling, and never-skilling—deserve consideration in the context of POCUS, each with different implications for training design and policy.

Deskilling occurs when clinicians have already achieved POCUS competency but experience progressive erosion of foundational skills because of sustained reliance on AI. Automation bias—the well-documented tendency to over-trust automated outputs—accelerates this trajectory [9]. When AI is consistently correct, vigilance erodes, and when AI fails, the clinician may lack the residual skill to catch the error. This can cause both omission and commission errors [4].

Misskilling refers to the adoption of errors or biases encoded within AI systems. POCUS AI tools that are trained predominantly on images from high-resource settings may perform poorly across diverse patient populations, body habitus, or disease prevalences. A clinician who calibrates their gestalt to AI outputs implicitly absorbs these algorithmic biases, potentially developing systematically flawed pattern recognition. This can be harder to detect and correct than simple skill loss.

Never-skilling is perhaps the most consequential risk on a generational timescale. Topaz and colleagues raise this concern directly: if trainees only learn with AI assistance, they may never develop the independent cognitive frameworks required for unassisted practice [4]. If educators themselves become reliant on AI systems, they lose the capacity to model expert reasoning and to teach trainees how to critically evaluate AI outputs, including identifying errors and hallucinations. In POCUS, current guidance recommends 150–300 total scans for baseline competency, with 25–50 per core application area [10]. If AI assistance is present throughout this formative exposure, trainees may complete their requirements without accumulating the independent perceptual experience that volume is intended to generate.

Lessons from Aviation and Other High-Stakes Industries

The Air France 447 disaster provides more than a cautionary metaphor; it illustrates a regulatory response that medicine has yet to replicate. Following that and similar incidents, aviation regulators mandated routine hands-on flight time and simulator programs designed to replicate automation failures, ensuring that the skills required for unexpected events remain intact [4]. For similar reasons, nuclear plant operators rehearse critical emergency scenarios at regular intervals: the consequence of deskilling in high-stakes moments is unacceptable.

These analogies point toward a principle that medicine applies in procedural simulation and cadaveric training but has not yet been systematically applied to POCUS in the context of AI assistance. The question is not whether AI should be integrated into POCUS practice; the evidence for its benefits in improving image quality, accelerating novice learning, and expanding access to bedside diagnostics is compelling [11, 12]. The question is how integration is designed to protect the human competencies that must persist when AI is unavailable, inaccurate, or insufficient.

A Framework for Preserving POCUS Competency

Protecting POCUS competency in an AI-augmented environment will require deliberate intervention at the level of workflow design, educational architecture, and institutional policy. Drawing on the five-point framework proposed by Topaz and colleagues [4] and adapting it to the POCUS context, we propose the following:

First, AI-off intervals should be built into clinical workflows. Analogous to aviation’s requirement for manual flight time, POCUS practice should incorporate deliberate periods without AI assistance; for example, one in every five scans performed without algorithmic support. In training settings, AI interpretation could be configured to appear only after the learner has logged independent findings, preserving the formative cognitive engagement that unassisted scanning demands. Institutions should track unaided performance metrics alongside AI-assisted ones to identify emerging skill gaps early.

Second, shadow-mode training should become standard in POCUS education. Learners should form independent assessments before seeing AI output, and simulation should expose trainees to AI failures and erroneous suggestions. This approach preserves active diagnostic reasoning while leveraging AI as a second reader rather than a primary one and ensures trainees can identify when algorithms are wrong.

Third, task boundary design offers a complementary approach for clinical practice. Frameworks should assign AI to high-volume, low-ambiguity sub-tasks confirming image adequacy, rather than applying AI uniformly across the POCUS task spectrum. AI can be used to flag unambiguous pathology while preserving the interpretive challenge of complex and equivocal cases for human clinicians. AI has demonstrated strong performance for detecting regional wall motion abnormalities in straightforward echocardiographic cases [13]. Such cases could be AI-triaged, which would free clinicians to focus on cognitive resources where nuanced human judgement is essential.

Fourth, competency-based assessment must evolve to explicitly account for AI exposure. Credentialing bodies should require demonstration of unaided POCUS competency as a distinct milestone separate from AI-assisted performance to ensure clinicians can function safely when technology fails [10]. As Topaz and colleagues argue, specialty boards should ensure that certification and maintenance assessments evaluate the ability to think critically and exercise independent clinical judgement, including competence in identifying AI errors [4].

Fifth, ongoing independent validation of AI tools must be a condition of clinical deployment. The COMPASS-AI survey of 1,154 healthcare professionals globally found that training and education (27.1%) and clinical validation and evidence (17.5%) were the most frequently cited barriers to adopting AI-use with POCUS [14]. Regulatory and professional bodies must move beyond requiring awareness of automation bias as the current European Union (EU) AI Act does to mandating active skill preservation, including post-market surveillance of unaided clinician performance [4].

Conclusion

AI is remaking the landscape of POCUS in ways that are genuinely promising. Tools that guide probe positioning, automate image quality assessment, and flag pathology in real time have the potential to expand access to bedside diagnostics and improve patient outcomes. Expert consensus has converged on the position that the goal must be augmentation, not replacement, of clinical skills [3].

Achieving that goal requires deliberate action. As Topaz and colleagues conclude, aviation learned through tragedy that automation paradoxically increases the importance of human vigilance rather than reducing it [4]. Health care and POCUS practice specifically is at a similar inflection point. The colonoscopy deskilling experience demonstrates that AI-related skill attrition is observable, measurable, and clinically significant [5]. In POCUS, where the skill set is multidimensional and the consequences of operator error are immediate, the risks of deskilling, misskilling, and never-skilling demand urgent and systematic attention. The choices made now about how AI is introduced to learners, how unaided competency is assessed, and how boundaries between human and algorithmic responsibility are drawn will determine whether AI strengthens POCUS practice or gradually erodes it from within. We can either heed these early warnings and build safeguards now, or risk discovering the costs of convenience only when systems fail and clinicians are no longer prepared to intervene.

Ethics Statement

This opinion editorial does not involve human participants, animal subjects, patient data, or any form of primary data collection. No ethical approval or informed consent was required for this work. All sources cited are publicly available, previously published literature.

Disclosure Statement

The author declares no conflicts of interest.

Funding

No funding was received for this work.

Author Contributions

AO: conceptualization, methodology, investigation, writing – original draft, writing – review & editing, project administration, visualization. TK: conceptualization, writing – original draft, writing – review & editing, validation. KF: conceptualization, writing – original draft, writing – review & editing, validation.

References

1. Kim J, Maranna S, Watson C, Parange N. A scoping review on the integration of artificial intelligence in point-of-care ultrasound: current clinical applications. Am J Emerg Med. 2025;92:172–181.

2. Towards Healthcare. AI in ultrasound imaging market size, share & growth analysis 2025–2034. 2025. Available from: https://www.towardshealthcare.com/insights/ai-in-ultrasound-imaging-market-size

3. Kumar A. AI in POCUS: expert panel reveals the future of point-of-care ultrasound. Global Ultrasound Institute. 2025. Available from: https://globalultrasoundinstitute.com/ai-pocus-future/

4. Topaz M, Peltonen LM, Zhang Z, Backonja U. The deskilling effect: is artificial intelligence eroding clinical competence? Ann Intern Med. 2026;179:1048–1050. doi: 10.7326/ANNALS-26-00613.

5. Budzyń K, Romańczyk M, Kitala D, Kołodziej P, Bugajski M, Adami HO, Blom J, Buszkiewicz M, Halvorsen N, Hassan C, Romańczyk T, Holme Ø, Jarus K, Fielding S, Kunar M, Pellise M, Pilonis N, Kamiński MF, Kalager M, Bretthauer M, Mori Y. Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study. Lancet Gastroenterol Hepatol. 2025;10(10):896–903. doi: 10.1016/S2468-1253(25)00133-5.

6. Dratsch T, Chen X, Rezazade Mehrizi M, Kloeckner R, Mähringer-Kunz A, Püsken M, Baeßler B, Sauer S, Maintz D, Pinto dos Santos D. Automation bias in mammography: the impact of artificial intelligence BI-RADS suggestions on reader performance. Radiology. 2023;307(4):e222176. doi: 10.1148/radiol.222176.

7. Qazi IA, Ali A, Khawaja AU, Akhtar MJ, Sheikh AZ, Alizai MH. Automation bias in large language model assisted diagnostic reasoning among AI-trained physicians: a randomized clinical trial. medRxiv. 2025. doi: 10.1101/2025.08.23.25334280.

8. Natali C, Marconi L, Dias Duran LD, Cabitza F. AI-induced deskilling in medicine: a mixed-method review and research agenda for healthcare and beyond. Artif Intell Rev. 2025;58(11):356. doi: 10.1007/s10462-025-11352-1.

9. Goddard K, Roudsari A, Wyatt JC. Exploring the risks of automation bias in healthcare artificial intelligence applications: a Bowtie analysis. Smart Health. 2024;33:100480.

10. ASRA News. POCUS Spotlight: defining competency in point of care ultrasound. American Society of Regional Anesthesia and Pain Medicine. 2025. Available from: https://asra.com

11. Shokoohi H, LeSaux MA, Roohani YH, Liteplo A, Huang C, Blaivas M. Acquisition of cardiac point-of-care ultrasound images with deep learning. CHEST Pulm. 2023. doi: 10.1016/j.chpulm.2023.100023.

12. Nhat PTH, Van Hao N, Tho PV, Kerdegari H, Pisani L, Thu LNM, Phuong LT, Duong HTH, Thuy DB, McBride A, Xochicale M, Schultz MJ, Razavi R, King AP, Thwaites L, Van Vinh Chau N, Yacoub S, Gomez A; VITAL Consortium. Clinical benefit of AI-assisted lung ultrasound in a resource-limited intensive care unit. Crit Care. 2023;27(1):257. doi: 10.1186/s13054-023-04548-w.

13. Slivnick JA, Gessert NT, Cotella JI, Oliveira L, Pezzotti N, Eslami P, Sadeghi A, Wehle S, Prabhu D, Waechter-Stehle I, Chaudhari AM, Szasz T, Lee L, Altenburg M, Saldana G, Randazzo M, DeCara JM, Addetia K, Mor-Avi V, Lang RM. Echocardiographic detection of regional wall motion abnormalities using artificial intelligence compared to human readers. J Am Soc Echocardiogr. 2024;37(7):655–663. doi: 10.1016/j.echo.2024.03.017.

14. Wong A, Roslan NL, McDonald R, Noor J, Hutchings S, D’Costa P, Via G, Corradi F. Clinical obstacles to machine-learning POCUS adoption and system-wide AI implementation (The COMPASS-AI survey). Ultrasound J. 2025;17(1):32. doi: 10.1186/s13089-025-00436-2.s13089-025-00436-2

Posted in .

Leave a Reply

Your email address will not be published. Required fields are marked *