Artificial Intelligence–Guided Lung Ultrasound by Nonexperts

Key Points

Question

Can artificial intelligence (AI) aid health care clinicians inexperienced in lung ultrasound (LUS) in obtaining high-quality LUS clips?

Findings

In this multicenter diagnostic validation study among adults with shortness of breath, 98.3% of ultrasound examinations performed by trained health care professionals with AI guidance were of sufficient quality to meet diagnostic standards and were not statistically different compared with images acquired by LUS experts without AI guidance.

Meaning

With AI assistance, trained novices can produce expert-level images that can be used to assess pathology after a short training session, potentially enhancing access to LUS in resource-constrained settings.

Abstract

Importance

Lung ultrasound (LUS) aids in the diagnosis of patients with dyspnea, including those with cardiogenic pulmonary edema, but requires technical proficiency for image acquisition. Previous research has demonstrated the effectiveness of artificial intelligence (AI) in guiding novice users to acquire high-quality cardiac ultrasound images, suggesting its potential for broader use in LUS.

Objective

To evaluate the ability of AI to guide acquisition of diagnostic-quality LUS images by trained health care professionals (THCPs).

Design, Setting, and Participants

In this multicenter diagnostic validation study conducted between July 2023 and December 2023, participants aged 21 years or older with shortness of breath recruited from 4 clinical sites underwent 2 ultrasound examinations: 1 examination by a THCP operator using Lung Guidance AI and the other by a trained LUS expert without AI. The THCPs (including medical assistants, respiratory therapists, and nurses) underwent standardized AI training for LUS acquisition before participation.

Interventions

Lung Guidance AI software uses deep learning algorithms guiding LUS image acquisition and B-line annotation. Using an 8-zone LUS protocol, the AI software automatically captures images of diagnostic quality.

Main Outcomes and Measures

The primary end point was the proportion of THCP-acquired examinations of diagnostic quality according to a panel of 5 masked expert LUS readers, who provided remote review and ground truth validation.

Results

The intention-to-treat analysis included 176 participants (81 female participants [46.0%]; mean [SD] age, 63 [14] years; mean [SD] body mass index, 31 [8]). Overall, 98.3% (95% CI, 95.1%-99.4%) of THCP-acquired studies were of diagnostic quality, with no statistically significant difference in quality compared to LUS expert–acquired studies (difference, 1.7%; 95% CI, −1.6% to 5.0%).

Conclusions and Relevance

In this multicenter validation study, THCPs with AI assistance achieved LUS images meeting diagnostic standards compared with LUS experts without AI. This technology could extend access to LUS to underserved areas lacking expert personnel.

Trial Registration

ClinicalTrials.gov Identifier: NCT05992324

This multicenter validation study evaluates the ability of artificial intelligence to guide acquisition of diagnostic-quality lung ultrasound images by trained health care professionals.

Introduction

Lung ultrasound (LUS) is increasingly used for diagnosis, assessment, and monitoring of patients with shortness of breath.¹ LUS is portable, is low cost, and does not expose patients to radiation, while allowing for rapid, real-time examination of the lungs for pathology.^1,2,3,4,5,6 Compared with chest radiography, LUS has better accuracy in detecting pneumonia, pneumothorax, pleural effusion, and pulmonary edema when used correctly.^7,8,9,10,11 In particular, the identification of B-line artifacts, along with additional point-of-care ultrasound assessments, such as echocardiography or evaluation of inferior vena cava diameter and collapsibility, can aid in the diagnosis and monitoring of conditions, including acute decompensated heart failure with pulmonary edema.^11,12,13,14 Evidence favors LUS use in emergency departments, intensive care units, and outpatient clinic settings across several medical specialties (eg, primary care, nephrology, heart failure clinics).^15,16,17,18

As opposed to other means of lung interrogation like chest radiography or computed tomography, which are less operator dependent, sonography requires hands-on training and adequate technical skill for diagnostic image acquisition.^19,20,21,22 Artificial intelligence (AI), particularly a subset of machine learning known as deep learning, has the potential to improve both the acquisition and interpretation of ultrasound images.²³ Deep learning relies on multilayered artificial neural networks to extract intricate patterns from high-dimensional data and has been particularly successful in computer vision tasks, such as image and video analysis. AI has been developed in cardiac ultrasound to guide acquisition of appropriate views, provide guidance to improve image quality, and automate assessments, such as estimating left ventricular ejection fraction.^24,25,26 AI may similarly be applied to LUS—however, the application of AI to LUS is still nascent.

Existing research has predominantly concentrated on development of AI algorithms for detecting and assessing the severity of ultrasound artifacts linked with diverse lung pathologies performed by expert users.^{27,28,29,30,31,32} However, AI solutions tailored to assist nonexpert users in acquiring LUS clips are needed. AI software focused on image acquisition is imperative for practical implementation to address user challenges in capturing diagnostic-quality clips and enable the widespread adoption of LUS across different health care settings, including among personnel without advanced ultrasound expertise. The purpose of this study was to evaluate the ability of AI technology to guide the acquisition of diagnostic-quality LUS images by trained health care professionals (THCPs) without significant LUS experience compared with expert LUS users.

Methods

Study Design

This prospective multicenter validation study was conducted between July 2023 and December 2023 across 4 geographically distinct clinical sites: 1 outpatient cardiovascular clinical site, 1 inpatient site, and 2 emergency departments. Participants were eligible if they were older than 21 years with clinical concern for pulmonary edema based upon symptoms or signs (eg, shortness of breath, hypoxia). Participants were enrolled consecutively and were excluded if they were in extremis, such that a research ultrasound could not be performed due to competing resuscitation priorities. Data on patient demographic characteristics, including age, sex, race, and ethnicity, were collected from the medical record at each institution and classified by trained staff using standardized categories defined by the investigators. Race and ethnicity were considered important to ensure a balanced dataset for AI validation.

After enrollment, participants underwent 2 LUS examinations. One was conducted by a THCP operator using the Lung Guidance AI software, and the other was conducted by either an ultrasound fellowship–trained physician (ie, an LUS expert) or by an advanced cardiac sonographer without the Lung Guidance AI. Details on the Lung Guidance AI background and software workflow are provided in eMethods 1 and 2 in Supplement 1. Examinations were conducted independently within 1 hour of each other, with operators not simultaneously present in the room and masked to each other’s findings. Each operator followed an 8-zone protocol for clip collection as described by Volpicelli and colleagues.³³ Experts captured clips manually, while THCP clips were captured using either the Save Best Clip or Autocapture features of the Lung Guidance software (eFigure 1 in Supplement 1). The Autocapture feature is an AI-driven component automating the clip capture when the AI determines that an adequate quality image has been obtained. If Autocapture failed to initialize, the operator had the option to press Save Best Clip, and the clip with the highest quality was saved (Figure 1). At least 1 clip per zone was saved prior to moving to each lung zone. In this study, a THCP was defined as an individual who underwent standardized training on the AI system used in the research and who used the software to capture clips (eMethods 3 in Supplement 1). Most THCPs were registered nurses and medical assistants who lacked any formal ultrasound training or prior clinical experience with diagnostic ultrasound (eTable in Supplement 1). However, a group of ultrasound-trained physicians was deliberately included as a subset within the THCP group to assess the impact of the software and autocapture on their ability to acquire studies meeting diagnostic standards. These 4 physicians had completed an emergency medicine residency program with prior LUS training. Our aim was to demonstrate that ultrasound-trained physicians using the software did not exhibit inferior performance. An a priori analysis comparing the results of physician THCPs to those of novice THCPs was planned. Nonphysician THCPs received 2.5 hours of supervised lung ultrasound practice with the software, while physician THCPs had no practice.

Figure 1. Overview of Model Development.

Overview of model development (left), key features (middle), and user interface with all features (right) for artificial intelligence (AI) Lung Guidance software.

All participants provided written informed consent to participate in the study. The study protocol was approved by the institutional review boards at each respective institution. This study was registered with ClinicalTrials.gov (NCT05992324) and adhered to the Standards for Reporting of Diagnostic Accuracy (STARD) reporting guidelines.³⁴

Lung Guidance Software

The investigational software provides guidance to facilitate LUS image acquisition and interpretation for B-lines annotation and autocapture (eMethods 4 and eFigure 2 in Supplement 1). Its core features are (1) a workflow feature that provides static guidance and organization of captured images based on an 8-zone LUS protocol; (2) a landmark-based guidance feature to assist the user to identify the correct landmarks for each lung zone; (3) a B-line annotation and autocapture feature to annotate and automatically capture images deemed to be of sufficient quality when B-lines are present; and (4) an 8-zone reporting feature to present the findings for each of the zones (Figure 1). The algorithms were interfaced with a commercially available ultrasound system (Butterfly IQ [Butterfly Network]). The main component of the AI that this study addressed is the Lung Guidance software, which uses the detection of landmarks in the lungs (eg, ribs, pleural line, A-line artifacts, diaphragm) to guide the user in performing a standard 8-zone protocol, consisting of upper anterior (zones 1 and 5), lower anterior (zones 2 and 6), upper lateral (zones 3 and 7), and lower lateral (zones 4 and 8). Detection of a clip of sufficient quality triggers the autocapture feature to record a clip 4 to 6 seconds in length.

Ground Truthing

A panel of 5 expert LUS readers independently reviewed all clips, with studies presented in a random order across patient and user type (THCP or LUS expert). The expert LUS readers were masked to both the study operators and the use of Lung Guidance AI, although the clips were labeled with the zone number. Readers underwent a priori training on the rating scheme. To determine diagnostic image quality, the expert reader panel evaluated if the clips were of sufficient quality to make 1 of the following LUS assessments: insufficient image information for each clip and each zone, normal examination, and abnormal examination. The panel assigned a binary score for diagnostic image quality, with the ground truth defined as the consensus (majority opinion) of the expert reader group. All LUS expert readers were physicians with completed fellowships in point-of-care ultrasound, leadership positions in national emergency medicine and/or ultrasound societies (eg, the Society for Academic Emergency Medicine Academy of Emergency Ultrasound, the American College of Emergency Physicians Emergency Ultrasound Section, and the American Institute of Ultrasound in Medicine), and published research in emergency and lung ultrasound.

Outcomes or End Points

The primary objective was to evaluate the Lung Guidance AI algorithm’s ability to help THCPs capture high-quality LUS clips that met diagnostic standards. This study focused on the acquisition of diagnostic-quality images, not on evaluating diagnostic accuracy or comparing nonexpert and expert operators’ diagnostic skills. The proportion of patient studies with sufficient quality to make a clinical assessment was assessed. Subgroup analyses were planned a priori based on age, sex, race and ethnicity, body mass index (BMI, calculated as weight in kilograms divided by height in meters squared), study site, and THCP level of expertise in LUS. Acquisition time using the Lung Guidance software was measured as a secondary end point.

Statistical Analysis

The Lung Guidance AI algorithm’s performance was considered successful if at least 80% of LUS clips met clinical assessment quality standards. This threshold was established based on a pilot study of 24 participants, where THCP-acquired images achieved a 94% patient-level point estimate (95% CI, 80.5%-99.3%). The 80% benchmark, derived from the lower bound of the confidence interval, aligns with standards in similar studies.²⁴ Power analysis indicated a minimum sample size of 130 patients to achieve 95% statistical power.

Statistical analyses included descriptive statistics and inferential analyses using 2-sided tests (α = .05). The proportion of sufficient quality examinations was evaluated using a χ² test of equal proportions, with 95% confidence intervals calculated using the Wilson score method.³⁵ All analyses were performed using R version 4.2.2 (R Foundation) and Python version 3.11.1 (Python Software Foundation).

Results

A total of 188 participants were enrolled across 4 centers, of whom 176 participants had a complete set of THCP examinations obtained with the use of Lung Guidance AI. This group was included in the analysis for the primary end point as the intention-to-treat population. A total of 163 participants had complete sets of data for both THCP examinations obtained with the use of Lung Guidance AI and expert examinations obtained without the use of Lung Guidance AI and were included in the THCP vs expert comparisons and the per-protocol analysis (Figure 2).

Figure 2. Enrollment Diagram.

THCP indicates trained health care professionals.

Of 176 participants with complete THCP examinations, 95 (54.0%) were enrolled in the emergency department sites and 80 (45.5%) in the other sites. A total of 81 participants (46.0%) were female, mean (SD) participant age was 63 (14) years, and the mean (SD) BMI was 31 (8) (Table 1). Across all participating institutions, a total of 21 THCPs took part in the study, composed of 3 medical assistants, 1 phlebotomist, 1 pharmacy technician, 12 registered nurses, and 4 physicians (eTable in Supplement 1). Registered nurses performed ultrasounds on 82 participants (46.6%), and physicians performed ultrasounds on 38 participants (21.6%) (eTable in Supplement 1).

Table 1. Participant Demographic Characteristics Based on 176 Participants With Complete Trained Health Care Professional Data.

Characteristic	Participants, No. (%)
Age, mean (SD), y	63 (14)
Sex
Female	81 (46.0)
Male	95 (54.0)
Racea
American Indian or Alaska Native	1 (0.6)
Black or African American	90 (51.1)
White	74 (42.1)
NA or not disclosedb	11 (6.3)
Ethnicitya
Hispanic or Latino	13 (7.4)
Not Hispanic or Latino	163 (92.6)
Body mass indexc
<25	46 (26.1)
≥25 to <30	52 (29.6)
≥30	78 (44.3)
Smoking status
Current	31 (17.6)
Never	59 (33.5)
Former	84 (47.7)
NA	2 (1.1)
Medical history
Asthma	32 (18.2)
Chronic obstructive pulmonary disease	41 (23.3)
Emphysema	7 (4.0)
Heart failure	124 (70.5)
Interstitial lung disease	3 (1.7)
Lung cancer	3 (1.7)
Pleural effusion	10 (5.7)
Pneumonia	13 (7.4)
Pulmonary embolism	13 (7.4)
Pulmonary hypertension	9 (5.1)
Suspected or confirmed COVID-19	7 (4.0)
Overall B-line prevalence	149 (84.7)

Abbreviation: NA, not available.

^a Data on patient demographic characteristics, including age, sex, race, and ethnicity, were collected from the medical record at each institution and classified by trained staff using standardized categories defined by the investigators.

^b Indicates participants whose medical records did not contain information regarding race or ethnicity or who did not disclose when asked.

^c Calculated as weight in kilograms divided by height in meters squared.

Overall, 98.3% of ultrasound studies (95% CI, 95.1%-99.4%) where THCPs acquired clips aided by the Lung Guidance AI were deemed of diagnostic quality by the expert panel. eFigure 3 in Supplement 1 shows representative clips acquired with AI by THCPs. There were no statistically significant differences in the study-level image quality between studies acquired by THCPs with the use of AI and those acquired by LUS experts without the use of AI for both intention-to-treat (difference, 1.7%; 95% CI, −1.6% to 5.0%; P = .31) and per-protocol analyses (difference, −1.2%; 95% CI, −2.9% to 0.5%; P = .16). THCP-acquired image quality was statistically comparable to the expert group at the zone level as well (Table 2). For zone 6 (left antero-inferior region) specifically, the THCP group outperformed the expert group (90.9% vs 77.3%; difference, 13.6%; 95% CI, 6.1%-21.1%; P < .001).

Table 2. Trained Health Care Professional (THCP)–Acquired Study-Level and Zone-Level Image Quality, Subgroup Analyses Among 176 Participants in the Intention-to-Treat Sample, and Comparison With Image Quality of Expert-Obtained Clips.

Parameter	Sample size, No.	Quality point estimate, % (95% CI)		P value
		THCP	Expert
Study-level image quality	176	98.3 (95.1-99.4)	96.6 (92.8-98.4)	.31
Age, y
<65	84	97.6 (91.7-99.3)	96.4 (90.0-98.8)	.65
≥65	92	98.9 (94.1-99.8)	96.8 (90.9-98.9)	.31
Body mass indexa
<25	46	100.0 (92.3-100.0)	93.5 (82.5-97.8)	.08
≥25 to <30	52	100.0 (93.1-100.0)	98.1 (89.9-99.7)	.32
≥30	78	96.2 (89.3-98.7)	97.4 (91.1-99.3)	.65
Sex
Female	81	97.5 (91.4-99.3)	96.3 (89.7-98.7)	.65
Male	95	99.0 (94.3-99.8)	96.8 (91.1-98.9)	.31
Site location
Site 1	77	97.4 (91.0-99.3)	100.0 (95.3-100.0)	.15
Site 2	47	100.0 (92.4-100.0)	93.6 (82.8-97.8)	.08
Site 3	34	97.1 (85.1-99.5)	91.2 (77.0-97.0)	.30
Site 4	18	100.0 (82.4-100.0)	100.0 (82.4-100.0)	>.99
Clinical setting
Heart failure clinic or inpatient	81	98.8 (93.3-99.8)	92.6 (84.8-96.6)	.05
Emergency department	95	97.9 (92.7-99.4)	100.0 (96.1-100.0)	.16
THCP profile
Physician	38	100.0 (90.8-100.0)	NA	NA
Registered nurse	82	97.6 (91.5-99.3)	NA	NA
Phlebotomist	19	100.0 (83.2-100.0)	NA	NA
Pharmacy technician	7	100.0 (64.6-100.0)	NA	NA
Medical assistant	30	96.7 (83.3-99.4)	NA	NA
Zone-level image quality
Zone 1	176	98.9 (96.0-99.7)	96.0 (92.0-98.1)	.09
Zone 2	176	95.5 (91.3-97.7)	95.5 (91.3-97.7)	>.99
Zone 3	176	95.5 (91.3-97.7)	97.2 (93.5-98.8)	.40
Zone 4	176	82.4 (76.1-87.3)	83.5 (77.3-88.3)	.78
Zone 5	176	96.6 (92.8-98.4)	96.0 (92.0-98.1)	.78
Zone 6b	176	90.9 (85.7-94.3)	77.3 (70.5-82.8)	<.001
Zone 7	176	96.0 (92.0-98.1)	93.2 (88.5-96.1)	.24
Zone 8	176	79.0 (72.4-84.4)	78.4 (71.8-83.9)	.90

Abbreviation: NA, not available.

^a Calculated as weight in kilograms divided by height in meters squared.

^b Statistically significant (P < .05).

For the a priori planned subgroup analyses based on the roles or professions of THCPs (physicians and nonphysician health care professionals), there was no statistically significant difference in overall examination-level image quality between LUS obtained by the 3 groups: nonphysician THCPs with Lung Guidance AI, physician THCPs with Lung Guidance AI, and LUS experts without the AI (Table 3). At the zone level, the image quality achieved by nonphysician THCPs was statistically equivalent to that of the expert group, except for zone 6, where the nonphysician THCP group exhibited superior performance compared to the expert group (Table 3). When comparing all 3 groups, significant differences in image quality were observed in zones 4, 6, and 8, with physician THCPs outperforming the other 2 groups (Table 3). A total of 2398 clips were obtained by the THCP group. Of these, 1937 clips (65.9%) were autocaptured. Zones 4 and 8 had the lowest rate of autocapture (32.1% and 29.2%, respectively). The mean (SD) time for an 8–lung zone examination with Lung Guidance AI was 16.5 (7.3) minutes, with a median acquisition time of 15.0 minutes. The mean (SD) acquisition time for the physician THCP group was 10.8 (4.2) minutes (median, 10.2 minutes), while the nonphysician THCP group had a mean (SD) acquisition time of 18 (7.2) minutes (median, 16.3 minutes).

Table 3. Nonphysician Trained Health Care Professional (THCP) Studies, Physician THCP–Acquired Studies, Zone-Level Image Quality, and Comparison With Image Quality of Expert-Obtained Clips.

Parameter	Quality point estimate, % (95% CI)			P value for comparison between groups
	Expert (n = 176)	Nonphysician THCP (n = 138)	Physician THCP (n = 38)
Study-level image quality	96.6 (92.8-98.4)	97.8 (93.8-99.3)	100 (90.8-100.0)	.45
Zone-level image quality
Zone 1	96.0 (92.0-98.1)	98.6 (94.9-99.6)	100.0 (90.8-100.0)	.21
Zone 2	95.5 (91.3-97.7)	94.2 (89.0-97.0)	100.0 (90.8-100.0)	.32
Zone 3	97.2 (93.5-98.8)	94.2 (89.0-97.0)	100.0 (90.8-100.0)	.17
Zone 4a	83.5 (77.3-88.3)	79.7 (72.2-85.6)	100.0 (90.8-100.0)	.01
Zone 5	96.0 (92.0-98.1)	96.4 (91.8-98.4)	100.0 (90.8-100.0)	.46
Zone 6a	77.3 (70.5-82.8)	89.9 (83.7-93.9)	100.0 (90.8-100.0)	<.001
Zone 7	93.2 (88.5-96.1)	95.7 (90.8-98.0)	100.0 (90.8-100.0)	.20
Zone 8a	78.4 (71.8-83.9)	75.4 (67.6-81.8)	100.0 (90.8-100.0)	.003

^a Statistically significant difference (P < .05).

For the THCP-acquired studies, the ground truth expert panel reached unanimous agreement in 84.1% of cases, while for the expert-acquired studies, the unanimous agreement rate was 81.8%. In all cases, at least 3 experts agreed on the final decision.

Discussion

When aided by AI guidance software, the THCP group acquired high-quality studies suitable for clinical assessment 98.3% of the time, surpassing the prespecified 80% benchmark. These results suggest that LUS users of various experience levels can obtain images of sufficient quality for diagnostic purposes when aided by an AI algorithm focused on image acquisition. Results were consistent across diverse patient demographics and BMIs. The Lung Guidance algorithm was tested with a diverse group of health care professionals, highlighting the algorithm’s versatility across different health care settings and in the hands of professionals with different levels of ultrasound experience. Additionally, the image quality produced by physician THCPs with AI guidance demonstrates that using the algorithm did not reduce the performance of ultrasound-trained physicians.

The median acquisition time for the AI-aided LUS examination was 15 minutes, longer than the 6-minute median reported by Patrick and colleagues³⁶ or the 8-minute median for first-year residents (operators with the least amount of training). While noteworthy, this difference may be inconsequential in resource-limited settings where the alternatives may be inability to otherwise perform LUS or reliance on other scarce or costly diagnostic imaging.

Strengths of this study include a diverse patient population, including a wide range of ages and BMIs, inclusion of patients from both clinic and emergency department settings, rigorous expert consensus for the ground truth, and direct comparison with images from LUS experts.

When analyzed by zone, the THCP group acquired high-quality images that were statistically comparable to the expert group in 7 of 8 zones and performed better than experts in zone 6. Prior research has suggested that zone 6 is particularly challenging to image due to the influence of the heart in this region, resulting in a more limited ability to capture these images.³⁷ These findings suggest that AI guidance can enhance visualization, allowing for a more comprehensive lung examination, particularly for lingular pneumonia.

The current scientific literature on AI for lung ultrasound primarily focuses on the development and validation of algorithms to identify artifacts, such as A-lines and B-lines, as well as pathological conditions like pleural effusion and consolidations, offering potential for large-scale data analysis, clinical decision-making, and enhanced prognosis.^{27,28,29,30,31,32,38} Other studies evaluated AI LUS algorithms in clinical contexts, primarily assessing interrater reliability for artifact identification rather than addressing image acquisition guidance. Russell and colleagues²² demonstrated fair to moderate interrater reliability for B-line quantification in heart failure patients, although their analysis was limited by small sample size and patients with low BMI. Nhat and colleagues³⁹ showed improved diagnostic performance with AI assistance among nonexpert clinicians, but focused primarily on A-lines in patients in dengue shock, limiting external validity.

Various approaches have been explored to improve LUS image acquisition and reduce operator dependence. Some studies have demonstrated the feasibility of patient self-performed 4-zone and 8-zone LUS after brief training, although this method is not suitable for all patients (eg, those with physical limitations).^40,41 Robotic systems for autonomous scanning and AI-assisted image capture offer another promising avenue, with the potential to complete scans in less than 4 minutes.^42,43 However, these systems are still in earlier stages of development and validation, are costly, and require specialized hardware. In contrast, AI-assisted image capture can be implemented on existing ultrasound machines through software updates, offering a potentially faster and more cost-effective solution that allows for human oversight during the scanning process.

This study adds to the literature by focusing on AI-guided image acquisition and the automatic capture of LUS clips. Identification and severity algorithms heavily depend on the availability of high-quality images, a requirement often challenging to meet in the absence of operators trained in LUS. The AI presented in this study addresses this deficiency. By enabling a wide spectrum of health care professionals to conduct LUS examinations, this AI solution has the potential to significantly enhance diagnostic accuracy and patient care.

Limitations

One limitation of this study is the lack of direct comparison between the image acquisition quality of THCP examinations conducted with AI assistance and those without it. Introducing a THCP examination without AI guidance into the protocol would have necessitated that participants be scanned 3 times, which would have been excessively burdensome to patients with symptomatic pulmonary edema. Additionally, the performance of a second THCP examination could have biased the THCPs, as operators would have had more practice with LUS.

THCPs may have improved during the study as a result of practice and AI software feedback, biasing in favor of the AI intervention. However, this is less likely, as scanning sessions were separated over time and there was insufficient repetition compared to recommended learning curves for LUS proficiency.⁴⁴ A 1-hour interval between assessments was deemed reasonable for stable, nonventilated participants. While potential fluid status changes due to undocumented diuretic use might have occurred, this likely had minimal impact on the primary outcome of image quality. However, such changes could have influenced the B-line–based autocapture feature’s triggering.

Another potential limitation concerns the inclusion of ultrasound-trained physicians in the THCP group, which might bias overall results in favor of the AI intervention. Nonetheless, this group is essential to consider as potential end users and aligns with the US Food and Drug Administration’s recommendations for good machine learning practices in medical device development, emphasizing that clinical study participants and datasets should represent the intended patient population. Given that physicians with LUS training might also use the AI, we deemed that assessing the performance of this group would be crucial. Subgroup analyses found that the nonphysician THCP group achieved similar image quality to the physician THCP group when compared with LUS experts.

The study population exhibited a high prevalence of cardiogenic B-lines, with 1 autocapture feature based on their presence. However, the primary outcome measure was image quality, irrespective of the pathology identified. This approach allows the findings to be potentially applicable to other LUS scenarios, such as evaluating pleural effusions and pulmonary consolidations.

Conclusions

In conclusion, THCPs aided by AI achieved comparable performance to expert LUS users in acquiring images meeting diagnostic standards following brief software-focused training. This technology has the potential to extend diagnostic capabilities to underserved areas lacking access to expert personnel. Future research aims to integrate AI algorithms for image guidance and autocapture with advanced capabilities for detecting B-lines, pleural effusions, consolidations, and pleural line abnormalities. Validation of integrated guidance and interpretation algorithms in clinical practice settings is essential to assess real-world effectiveness and usability.

Supplement 1.

eMethods 1. Background

eMethods 2. Workflow Description

eMethods 3. THCP Training Details

eMethods 4. Model Development

eTable. Trained Health Care Provider Scan Distribution

eFigure 1. Workflow for Image Capture in Each Zone

eFigure 2. Overview of Model Development (Left), Key Features (Middle), and User Interface With All Features (Right) for Caption AI Lung Guidance Software

eFigure 3. Representative Diagnostic Quality Ultrasound Lung Clips Acquired With the Caption Lung AI by Lung Ultrasound Novice THCP

eReferences.

Supplement 2.

Data Sharing Statement