Letter to the Editor: Artificial intelligence in nephrology point-of-care ultrasonography - opportunities, limitations, and future directions

Abstract

We read with interest the retrospective cohort study by Silipigni et al entitled “Role of point-of-care ultrasonography in kidney disease management: A single solution for multiple challenges”. Artificial intelligence (AI) is rapidly reshaping the way clinicians learn, perform, and interpret point-of-care ultrasonography (POCUS). In nephrology, AI-assisted image acquisition, automated measurements, and emerging decision-support tools offer opportunities to improve efficiency and expand access to POCUS training. At the same time, important limitations remain, particularly when ultrasound findings must be integrated with physiology and clinical context. In this commentary, we discuss the current role of AI in nephrology POCUS, examine potential pitfalls of overreliance on automated interpretation, and consider future applications that may extend beyond image analysis toward physiologic assessment and hemodynamic phenotyping. Ultimately, we suggest that AI should be used to enhance, not replace clinicians’ understanding of anatomy, physiology, and patient-specific pathophysiology.

Key Words: Point-of-care ultrasound; Nephrology; Artificial intelligence; Automation; Large language models

Core Tip: Artificial intelligence (AI) has the potential to expand access to nephrology point-of-care ultrasonography through guidance of image acquisition, measurement automation, quality assurance, and clinical decision support. However, accurate image recognition does not necessarily equate to accurate physiologic interpretation, particularly in multi-organ hemodynamic assessment. AI should therefore be viewed as a tool that augments clinical judgment rather than replaces it.

TO THE EDITOR

We thank Silipigni and team for their thoughtful commentary on our review, “Point-of-care ultrasonography in nephrology: Growing applications, misconceptions and future outlook”[1]. In their article, “Role of point-of-care ultrasonography in kidney disease management: A single solution for multiple challenges”, the authors provide additional perspectives on the expanding role of point-of-care ultrasonography (POCUS) and training considerations[2]. Among several future directions discussed, the authors briefly acknowledge the emerging role of artificial intelligence (AI) in POCUS. Given the rapidly increasing interest in AI-assisted imaging and the unique educational and operational challenges being faced by nephrology POCUS programs, we would like to further expand this discussion by examining the opportunities, limitations, and future implications of AI in this space.

In nephrology, multiple surveys have consistently identified a lack of trained faculty, limited access to mentorship, and insufficient opportunities for longitudinal supervision as major barriers to POCUS implementation[3,4]. While educational initiatives continue to expand, the demand for ultrasound training substantially exceeds the availability of experienced instructors. In this setting, AI has emerged as a potentially valuable tool to augment training, standardize image acquisition, improve workflow efficiency, and support quality assurance. However, as with any emerging technology, its value depends not only on what it can do, but also on how it is used and integrated into clinical practice.

CURRENT APPLICATIONS OF ARTIFICIAL INTELLIGENCE IN NEPHROLOGY POCUS

Current AI applications in POCUS can broadly be divided into image acquisition assistance and automated image analysis. During image acquisition, AI-based transducer guidance systems provide real-time feedback regarding transducer movement and orientation, while automated structure recognition and labeling assist novice users in identifying relevant anatomy. These tools may reduce the initial intimidation often experienced by beginners after attending introductory workshops and may encourage earlier adoption of POCUS into clinical practice.

However, the benefits of AI-guided acquisition must be balanced against the risk of overreliance. For example, in cardiac POCUS, a cornerstone of hemodynamic assessment, understanding the three-dimensional orientation of cardiac anatomy is essential for acquiring diagnostically meaningful images. Unlike many abdominal applications where organ orientation is relatively intuitive, cardiac ultrasound requires a conceptual understanding of anatomy to accurately direct the ultrasound beam and obtain high-quality diagnostic views. If users become overly dependent on automated labeling and guidance without developing this foundational understanding, image acquisition may remain challenging whenever AI guidance is unavailable, or more importantly, inaccurate. Furthermore, failure to develop a mental three-dimensional model may hinder long-term skill acquisition and limit the ability to adapt when standard imaging windows are inaccessible or anatomy is altered by disease or prior surgery. These concerns align with the broader concept of AI-induced “never-skilling”, whereby excessive dependence on AI during formative learning may impede the development of fundamental cognitive and technical competencies required for independent practice[5]. AI should therefore serve as a bridge to learning, not a substitute for it.

Beyond acquisition guidance, AI has demonstrated substantial utility in automating measurements and reducing workflow burden[6]. Automated ejection fraction estimation, left ventricular outflow tract velocity-time integral (LVOT VTI) tracing, inferior vena cava (IVC) measurements, pulmonary B-line quantification and summaries, and report generation can significantly improve efficiency (Figure 1). These applications may be particularly valuable for experienced operators because they reduce repetitive tasks, facilitate documentation, shorten examination time, and potentially improve consistency. AI-assisted quality assurance systems may also help identify studies that meet predefined acquisition standards in institutions where expert POCUS reviewers are unavailable, thereby supporting educational programs and billing compliance.

Figure 1 Representative artificial intelligence assisted ultrasound applications in nephrology point-of-care ultrasonography. A: Automated left ventricular outflow tract velocity-time integral (LVOT VTI) tracing on the GE Venue system (GE HealthCare, Waukesha, WI, United States). Low sweep-speed acquisition as shown here may increase susceptibility to erroneous measurements; B: Smart inferior vena cava (IVC) analysis on the TE7 Max system (Mindray North America, Mahwah, NJ, United States) with automated tracking of longitudinal IVC motion; however, cardiac pulsatility may be incorporated into respiratory variation measurements; C: Automated lung ultrasound B-line summary on the Lumify handheld ultrasound system (Philips Ultrasound, Cambridge, MA, United States). Because each lung zone includes multiple intercostal spaces, serial comparisons may be affected by differences in probe position during follow-up examinations. Dmin: Minimum diameter of the IVC; Dmax: Maximum diameter of the inferior vena cava; DI: Distensibility index; IVCV: Inferior vena cava diameter variation.

Nevertheless, automated measurements remain highly dependent on underlying image quality and acquisition technique. For example, an AI algorithm may accurately trace a Doppler envelope while failing to recognize that the underlying apical view is tilted and Doppler alignment is suboptimal. Consequently, the reported LVOT VTI may appear precise while remaining fundamentally inaccurate. Similarly, automated measurements may be influenced by technical factors such as sweep speed selection. Lower sweep speeds allow averaging of multiple cardiac cycles, which is often desirable, but may also increase susceptibility to tracing artifacts or inclusion of nonrepresentative Doppler signals. Users must therefore understand not only how measurements are generated but also when AI-generated values are clinically meaningful.

Similar challenges exist in lung POCUS. Automated B-line quantification systems can improve efficiency and documentation but may occasionally misclassify shorter reverberation artifacts as true B-lines and lead to overcounting[7]. Furthermore, AI-generated summaries may create a false impression of precision during serial examinations because commonly used lung ultrasound protocols assess broad thoracic zones rather than defined intercostal spaces. Small variations in probe position between examinations may influence B-line counts and affect longitudinal monitoring. As with all automated measurements, clinician review remains essential.

Automated IVC assessment provides another illustrative example. Modern AI algorithms can track longitudinal vessel motion and address an important limitation of traditional M-mode measurements, wherein the IVC may move out of the sampling cursor during respiration. This represents a meaningful technological advance. However, challenges remain. AI systems may incorrectly attribute cardiac pulsatility to respiratory variation when calculating collapsibility indices or fail to recognize technical pitfalls such as the cylinder effect (spurious diameter variation from off-axis imaging of a three-dimensional vessel). More importantly, even when measurements are technically accurate, AI-derived summaries or interpretations may be misleading if they do not account for physiologic and pathologic factors that fundamentally alter IVC behavior, including mechanical ventilation, elevated intra-abdominal pressure, severe tricuspid regurgitation, or pulmonary hypertension.

CURRENT LIMITATIONS AND FUTURE DIRECTIONS

While image acquisition and measurement automation have received considerable attention, the greatest future challenge for AI in nephrology POCUS may lie in clinical integration. The primary value of POCUS is not image acquisition itself but the synthesis of imaging findings into clinically meaningful physiologic assessments. This distinction is particularly important in nephrology, where assessment of hemodynamics increasingly relies on multi-organ ultrasound rather than isolated findings.

For example, a dilated IVC should not automatically trigger conclusions regarding hypervolemia, nor should a small collapsible IVC be equated with hypovolemia. Rather, the IVC is best viewed as a surrogate of right atrial pressure, which is determined by multiple factors beyond volume status and should not be interpreted in isolation. Similarly, abnormal venous Doppler findings used to assess systemic venous congestion require integration with cardiac function, respiratory mechanics, cardiac rhythm, local structural changes affecting the organ or vessel being interrogated (e.g., renal interstitial edema or cirrhosis), and the overall hemodynamic profile. An algorithm may accurately identify waveform patterns yet be unable to determine the underlying physiologic mechanisms responsible for those findings. This distinction highlights an important limitation of many contemporary AI systems: They excel at “pattern recognition” but remain far less capable of physiologic reasoning.

An additional and rapidly evolving area involves the use of large language models (LLMs). Unlike traditional computer vision algorithms that primarily focus on image recognition and measurement automation, LLMs may facilitate interpretation, reporting, education, and clinical decision support. Potential applications include automated generation of structured ultrasound reports, translation of technical findings into clinically meaningful summaries[8], educational feedback for trainees, and synthesis of ultrasound findings with laboratory data, hemodynamic variables, and other clinical information. Such capabilities may be particularly attractive in nephrology, where POCUS findings are often interpreted alongside complex physiologic and biochemical data. However, important limitations remain. LLMs do not inherently understand physiology and may generate plausible but incorrect explanations when provided with incomplete, inaccurate, or poorly contextualized inputs. Consequently, the quality of their output remains highly dependent on the quality of the information provided. Rather than functioning as independent interpreters, LLMs are more appropriately viewed as tools that may augment clinician workflow, communication, and reasoning while remaining subject to expert oversight.

The next frontier for AI in nephrology POCUS will likely extend beyond image recognition and measurement automation toward integrated clinical decision support. Custom multimodal LLMs coupled with retrieval-augmented generation may also help address one of the major challenges in POCUS education: Clinical interpretation. By integrating ultrasound images with relevant clinical data and curated educational resources, these systems could provide context-specific explanations and differential diagnoses tailored to the learner. Beyond education, combining ultrasound findings with other clinical data sources may facilitate physiologic assessment and hemodynamic phenotyping, although rigorous validation will be essential before such approaches can be widely adopted.

CONCLUSION

AI has the potential to expand access to nephrology POCUS, reduce technical barriers to adoption, and improve workflow efficiency. However, its greatest value will likely be as an adjunct that enhances clinician expertise through acquisition guidance, measurement automation, quality assurance, education, and clinical decision support rather than as a replacement for clinical reasoning. As nephrology POCUS continues to evolve toward comprehensive multi-organ and hemodynamic assessment, successful integration of AI will require a continued emphasis on physiology, competency-based training, and thoughtful clinical implementation. Ultimately, the promise of both POCUS and AI lies in helping clinicians better understand the pathophysiologic state of individual patients and apply that understanding to improve patient care.