Abstract
Background
This trial evaluated whether an artificial intelligence (AI)‐based automatic analysis for echocardiography could improve sonographer workflow in real‐world clinical practice.
Methods
In a single‐center crossover trial, 4 sonographers were randomly assigned to use AI assistance (AI days) or manual workflow (non‐AI days) on a daily basis. The AI tool automatically measured echocardiographic parameters, allowing sonographers to focus on verifying AI‐generated values. Expert echocardiologists finalized all reports. The primary end point was examination efficiency, measured by examination time per patient and number of examinations per day. Secondary end points included sonographer fatigue, the number of analyzed echocardiographic parameters, and image quality.
Results
A total of 585 patients were scanned over 38 study days (AI days, 317; non‐AI days, 268) between January 30 and March 26, 2024. Baseline characteristics were comparable between groups. AI assistance significantly reduced examination time (13.0±3.5 minutes versus 14.3±4.2 minutes, P<0.001) and increased the number of daily examinations (16.7±2.5 versus 14.1±2.5, P=0.003). The number of echocardiographic parameters per examination increased 3.4‐fold on AI days (85±12 versus 25±1, P<0.001). Despite the higher workload, sonographers reported lower mental fatigue scores on AI days (4.1±1.1 versus 4.7±0.6, P=0.039). AI‐generated measurements agreed with final expert‐endorsed values within acceptable clinical limits for 90% of parameters. Notably, image quality significantly improved on AI days (P<0.001).
Conclusions
This real‐world randomized trial demonstrated that AI‐based echocardiographic analysis enhances workflow efficiency, reduces sonographer fatigue, and improves image quality without compromising diagnostic integrity. AI integration holds promise for optimizing high‐volume echocardiography workflows.
Registration
URL: https://center6.umin.ac.jp. Unique identifier: UMIN000053259.
Keywords: artificial intelligence, automated analysis, echocardiography, randomized trial, work burden
Subject Categories: Echocardiography, Ultrasound, Digital Health
Nonstandard Abbreviations and Acronyms
- EchoNet‐RCT
Blinded, Randomized Controlled Trial of Sonographer Versus Artificial Intelligence Assessment of Cardiac Function
Clinical Perspective.
What Is New?
This randomized crossover trial is the first to evaluate daily use of artificial intelligence–based automated echocardiographic analysis in real‐world clinical practice, demonstrating reduced examination time, increased daily scan volume, and improved image quality.
What Are the Clinical Implications?
Artificial intelligence integration enabled sonographers to complete more echocardiograms with less mental fatigue, supporting its role in sustaining high‐volume laboratories while maintaining diagnostic quality.
By enhancing efficiency and image quality, artificial intelligence–assisted workflows may facilitate broader use of advanced echocardiographic parameters and improve patient care in routine cardiovascular practice.
Echocardiography is a cornerstone, noninvasive diagnostic tool in cardiovascular medicine, widely used for its ability to provide detailed insights into cardiac structure and function. Over recent years, the increasing prevalence of cardiovascular disease has driven a corresponding rise in the demand for echocardiographic examinations. 1 This trend has been particularly pronounced in the context of screening echocardiography, which often involves repetitive, standardized tasks. Such a large number of examinations in routine workflows can diminish sonographers’ motivation, contributing to fatigue and elevated rates of burnout. 2 Addressing this issue, optimizing the efficiency of echocardiographic workflows has become an urgent priority, especially in countries like Japan, where the number of echocardiographic examinations per capita is 3.5 times higher than in the United States. 3
In recent years, artificial intelligence (AI) has advanced rapidly, with significant potential applications in health care. 4 , 5 , 6 In echocardiography, AI excels in image recognition and has shown great promise in the automatic classification and analysis of echocardiographic images. 7 , 8 Recent studies have validated AI’s capability to automate the analysis of most routine echocardiographic parameters, reducing analysis time while maintaining high reproducibility. 9 , 10 , 11 , 12 These tools offer the potential to streamline sonographers’ workflows by reducing time spent on routine measurements, thereby allowing greater focus on more complex tasks, such as managing challenging patient cases or interpreting echocardiographic findings. 13 However, much of the existing research has been conducted retrospectively or in controlled experimental environments, leaving questions about AI’s impact on workflow efficiency in real‐world clinical settings.
To address this gap, we conducted a randomized crossover trial to evaluate whether an AI‐based automatic analysis tool for echocardiography could streamline the daily workflow of sonographers in a real‐world clinical environment.
Methods
Data Availability
The data underlying this article cannot be shared publicly for the reason of maintaining the privacy of individuals who participated in the study. The data will be shared upon reasonable request to the corresponding author.
Study Design
This single‐center, single‐blinded, randomized crossover study was conducted at Juntendo University Hospital (Tokyo, Japan) between January 2024 and March 2024. Certified clinical sonographers, credentialed by the Japan Society of Ultrasonics in Medicine, were enrolled in the study after providing written informed consent. The study protocol received approval from the Research Ethics Committee of the Faculty of Medicine, Juntendo University (approval number: E23‐0161‐M01), and the study was registered in the University Hospital Medical Information Network Clinical Trials Registry (identification number: UMIN000053259) before participant enrollment. All procedures adhered to the principles outlined in the Declaration of Helsinki, and the Consolidated Standards of Reporting Trials checklist 14 is available in Table S1.
Randomization of AI and Non‐AI Days
The study used a randomized crossover design, wherein echocardiographic examinations were conducted either with AI assistance (AI days) or without AI assistance (non‐AI days). The echocardiography laboratory comprised 8 booths, each staffed by a sonographer on a daily basis. One specific booth with a dedicated ultrasound machine (Epic Elite; Philips Healthcare, Netherlands) was designated for the randomized AI or non‐AI workflow throughout the study, while the remaining booths followed standard procedures. The allocation of AI and non‐AI days for this designated booth was determined using computer‐generated block randomization, considering the day of the week. A custom software program was used to manage the randomization, and sonographers accessed the program each morning to determine whether that day’s examinations would involve AI assistance.
The laboratory handled 2 types of examinations: “detailed” echocardiography for patients with known cardiac diseases, and “screening” echocardiography for preoperative assessments, chest pain evaluations, and for patients without a known history of cardiovascular disease. However, the study booth focused exclusively on screening examinations, as these repetitive and formulaic tasks are precisely the type of examinations that AI is expected to streamline.
AI‐Assisted Workflow (AI Days)
On AI days, sonographers used an automated AI analysis tool, specifically the Food and Drug Administration–approved US2.ai platform (Us2.ai, Singapore). This platform automatically measured key echocardiographic parameters (detailed in Table S2), eliminating the need for manual measurements in most cases. 11 Images were uploaded to an on‐premise server, and AI analysis results were sent back to the reporting system (ISCV; Philips Healthcare, Netherlands) within ≈2 minutes. Sonographers then reviewed these AI‐generated results, which included trace lines overlaid on the images to illustrate the AI analysis (Figure S1). Certain parameters, such as the diameters of the left atrium, aortic root, and inferior vena cava, were not measured by AI and required manual measurement. If any AI‐generated measurements or tracings were considered inaccurate, the sonographer remeasured the parameters manually on the workstation. Both AI and non‐AI reports were reviewed and finalized by expert echocardiologists, who were blinded to the AI status, ensuring all reports were suitable for clinical use.
Standard Workflow (Non‐AI Days)
On non‐AI days, sonographers performed echocardiographic examinations without AI assistance, following their routine clinical procedures. Sonographers were allowed to select their preferred method for conducting measurements, whether directly on the echocardiography machine or at the workstation after the scan. All measurements adhered to established guidelines. 15 , 16
Study End Points
The primary end points of this study were the efficiency of echocardiographic examinations, measured by the total time required for scanning, analysis, and reporting for each patient, as well as the total number of patients scanned and reported on by each sonographer during a typical workday (from 9:00 am to 5:00 pm).
Secondary end points included (1) sonographers’ mental and physical fatigue at the end of each workday, measured using a daily questionnaire shown in Table S3; (2) the number of echocardiographic parameters analyzed per examination; (3) the quality of echocardiographic images; and (4) AI’s performance, assessed by the rate at which AI analyzed acquired images and the concordance between AI’s initial measurements and the final values endorsed by expert echocardiologists.
The image quality was evaluated by 2 blinded echocardiologists who assessed 5 standard views: parasternal long‐axis, short‐axis, apical 4‐chamber, 3‐chamber, and 2‐chamber views. Each view was graded on a 3‐point scale (poor, >5 segments poorly visible; good, 3–5 segments poorly visible; and excellent, only 0–2 segments poorly visible throughout the cardiac cycle), adapted from a previous report. 9 , 17
Sample Size Calculation
The required number of study days was determined on the basis of a sample size calculation. Previous research has demonstrated that using this AI system results in a 70% reduction in analysis time. 9 However, considering the complexity of real‐world processes in an echocardiographic laboratory, we conservatively estimated an overall efficiency improvement of 20%, corresponding to an increase in the number of examinations per day from ≈10±2 to 12±2. The sample size calculation indicated that 16.7 days per arm would be necessary to detect this difference (α=0.05, β=0.8). To account for an estimated 10% data attrition, we ultimately set the study to include 19 days of examinations per arm.
Statistical Analysis
Data are presented as mean±SD or median (interquartile range) for continuous variables, as appropriate, and as frequency (percentage) for categorical variables. Group differences were evaluated using Welch’s t test or the Mann–Whitney U test for continuous variables and the χ2 test or Fisher’s exact test for categorical variables. Bland–Altman plots were created to visualize systematic errors between the AI initial measurements and the final report values, with limits of agreement defined by the mean bias±1.96 SDs. Interrater reproducibility of the image quality assessments between the 2 echocardiologists was evaluated using intraclass correlation coefficients (1, 2). All statistical analyses were performed using R version 4.4.1 (R Foundation for Statistical Computing, Vienna, Austria). For all analyses, a 2‐tailed P value <0.05 was considered statistically significant.
Results
Sonographers and Examination Characteristics
During the study period, 4 sonographers (average experience in echocardiography 9.0±4.4 years, all women) participated in the study, scanning a total of 585 patients over 38 days, with 19 days each allocated to AI days and non‐AI days. These sonographers had experienced the AI‐based workflow for 1 month before starting the study. As summarized in Table 1, the characteristics recorded in the sonographers’ reports, such as patient sex, age, body size, and echocardiographic parameters, were well balanced between AI and non‐AI days. Most scans were performed on patients in sinus rhythm, with <4% showing atrial fibrillation. The echocardiographic parameters remained largely within normal ranges, 18 , 19 and the prevalence of valvular heart disease was minimal, reflecting that the examinations were for screening purpose.
Table 1.
Scan Characteristics on AI and Non‐AI Days
| Non‐AI day (N=268) | AI day (N=317) | |
|---|---|---|
| Women, n (%) | 144 (53.7) | 191 (60.3) |
| Age, y, mean±SD | 64±16 | 65±15 |
| Body mass index, mean±SD | 22.9±3.8 | 23.2±4.4 |
| Body surface area, m2, mean±SD | 1.63±0.19 | 1.62±0.21 |
| ECG, n (%) | ||
| Sinus rhythm | 250 (93.3) | 311 (98.1) |
| Atrial fibrillation | 10 (3.7) | 4 (1.3) |
| Others | 8 (3.0) | 2 (0.6) |
| LVIDd, mm, mean±SD | 44±5 | 44±5 |
| LVIDs, mm, mean±SD | 29±4 | 28±5 |
| IVSTd, mm, mean±SD | 9±2 | 9±2 |
| LVEF (2‐dimensional disk), %, mean±SD | 63±8 | 64±8 |
| Left atrial diameter, mm, mean±SD | 34±7 | 34±6 |
| E/A, mean±SD | 0.97±0.40 | 0.97±0.43 |
| Aortic stenosis grade, n (%) | ||
| After aortic valve surgery | 2 (0.7) | 2 (0.6) |
| None/trivial | 260 (97.0) | 307 (96.8) |
| Mild | 5 (1.9) | 4 (1.3) |
| Moderate/severe | 1 (0.4) | 4 (1.3) |
| Aortic regurgitation grade, n (%) | ||
| After aortic valve surgery | 2 (0.7) | 2 (0.6) |
| None/trivial | 226 (84.3) | 282 (89.0) |
| Mild | 28 (10.4) | 25 (7.9) |
| Moderate/severe | 12 (4.5) | 8 (2.5) |
| Mitral regurgitation grade, n (%) | ||
| After mitral valve surgery | 2 (0.7) | 3 (0.9) |
| None/trivial | 240 (89.6) | 283 (89.3) |
| Mild | 22 (8.3) | 26 (8.3) |
| Moderate/severe | 4 (1.5) | 5 (1.6) |
| Tricuspid regurgitation grade, n (%) | ||
| None/trivial | 222 (82.8) | 261 (82.3) |
| Mild | 36 (13.4) | 44 (13.9) |
| Moderate/severe | 10 (3.7) | 12 (3.8) |
AI indicates artificial intelligence; E/A, early to late ventricular filling velocity ratio; IVSTd, interventricular septal thickness in diastole; LVEF, left ventricular ejection fraction; LVIDd, left ventricular internal diameter in diastole; and LVIDs, left ventricular internal diameter in systole.
Examination Efficiency and Sonographer Workload
The primary end point of average examination time was reduced to 13.0±3.5 minutes on AI days, compared with 14.3±4.2 minutes on non‐AI days (P<0.001; Figure 1A). As a result, the number of echocardiography scans performed per sonographer increased significantly on AI days, averaging 16.7±2.5 examinations per day, compared with 14.1±2.5 on non‐AI days (P=0.003; Figure 1B). Over the course of the study, 317 examinations were performed on AI days and 268 on non‐AI days. Additionally, the number of echocardiographic parameters analyzed per examination increased 3.4‐fold on AI days compared with non‐AI days (85±12 versus 25±1, P<0.001; Figure 2A). Notably, left ventricular (LV) and atrial strain were analyzed on 90% and 80% of AI days, respectively, whereas these parameters were never analyzed on non‐AI days.
Figure 1. Primary end points of examination efficiency.
On AI days, the average examination time was reduced compared with non‐AI days (A). As a result, the number of echocardiograms performed per sonographer increased significantly on AI days (B). AI indicates artificial intelligence.
Figure 2. Number of parameters and image quality.
On AI days, the number of analyzed parameters increased by up to 3.4‐fold compared with non‐AI days (A). Additionally, image quality was significantly higher on AI days (B). AI indicates artificial intelligence.
Despite the increase in the number of examinations, the Likert scale–based questionnaire results indicated that mental fatigue experienced by sonographers was significantly lower on AI days compared with non‐AI days (4.1±1.1 versus 4.7±0.6, P=0.039). The questionnaire also revealed that sonographers’ physical fatigue (4.0±0.9 versus 4.5±0.8, P=0.088) and perception of task complexity was numerically lower on AI days (3.7±1.0 versus 4.2±0.8, P=0.21), and that their perceived work speed was numerically faster on AI days (3.9±1.1 versus 3.3±1.6, P=0.26).
Improved Image Quality on AI Days
We compared the quality of images acquired on AI days with those obtained on non‐AI days to test whether, on AI‐assisted days, sonographers could dedicate more attention to image acquisition, resulting in higher‐quality echocardiographic images. Two experienced echocardiologists, who were not involved in the report finalization and were blinded to the examination allocation and study results, independently reviewed all images using the evaluation methods outlined in the Methods section. Interobserver variability tested on 150 images showed excellent agreement, with an intraclass correlation coefficient of 0.82. A total of 2925 images were evaluated and categorized as poor, good, or excellent quality. The results, shown in Figure 2B, demonstrated that percentage of images rated as excellent was significantly higher on AI days, with 41% of images achieving this rating compared with 31% on non‐AI days (P<0.001).
AI’s Performance in Real‐World Clinical Practice
To assess AI’s performance in this trial, the rate at which AI analyzed acquired images and the concordance between AI’s initial measurements and the final values endorsed by expert echocardiologists were analyzed, as summarized in Table 2. Overall, AI successfully analyzed most parameters in ≈95% of the cases from available images, while transmitral inflow parameters and left atrial volume index showed an analysis rate of 80% to 85%.
Table 2.
Concordance Between AI’s Initial Measurements and Final Report Values
| Parameters | AI’s initial measurements, mean±SD | Final report values, mean±SD | Available images, N | Images analyzed by AI, n (%) | Acceptable rate of AI values,* % | Mean absolute modification by sonographer† | ICC between AI and final values | P values for ICC |
|---|---|---|---|---|---|---|---|---|
| IVSTd, mm | 9.5±2.0 | 9.5±1.6 | 316 | 313 (99.1) | 94.6 | 1.7 | 0.81 | <0.001 |
| PWTd, mm | 9.3±1.8 | 9.3±1.6 | 316 | 310 (98.1) | 94.6 | 1.7 | 0.80 | <0.001 |
| LVIDd, mm | 42.6±5.6 | 43.6±4.8 | 316 | 314 (99.4) | 86.7 | 3.4 | 0.84 | <0.001 |
| LVIDs, mm | 28.5±6.3 | 28.1±4.6 | 316 | 311 (98.4) | 70.6 | 3.9 | 0.76 | <0.001 |
| LVEDV, mL | 71.1±22.2 | 72.9±24.4 | 299 | 279 (93.3) | 94.3 | 8 | 0.95 | <0.001 |
| LVESV, mL | 25.9±12.5 | 26.7±13.9 | 299 | 279 (93.3) | 92.0 | 5.2 | 0.95 | <0.001 |
| LVEF (2‐dimensional disk), % | 64.3±8.3 | 64.5±8.0 | 299 | 279 (93.3) | 91.3 | 3.9 | 0.92 | <0.001 |
| MV‐E, cm/s | 77.3±21.3 | 74.5±20.6 | 314 | 262 (83.4) | 95.2 | 4.8 | 0.96 | <0.001 |
| MV‐A, cm/s | 84.6±21.0 | 82.1±23.3 | 298 | 249 (83.6) | 96.3 | 3.1 | 0.99 | <0.001 |
| E/A | 0.97±0.43 | 0.94±0.33 | 298 | 244 (81.9) | 99.0 | 0.1 | 0.98 | <0.001 |
| Deceleration time, ms | 213±53 | 215±38 | 314 | 228 (72.6) | 90.4 | 26 | 0.78 | <0.001 |
| e’ (septal), cm/s | 8.4±2.6 | 8.5±2.5 | 316 | 306 (96.8) | 94.9 | 0.5 | 0.98 | <0.001 |
| e’ (lateral), cm/s | 10.3±3.4 | 10.3±3.1 | 317 | 313 (98.7) | 94.0 | 0.6 | 0.99 | <0.001 |
| E/e’ (septal) | 9.5±3.3 | 9.8±3.7 | 316 | 255 (80.7) | 91.8 | 0.6 | 0.92 | <0.001 |
| E/e’ (lateral) | 7.7±2.8 | 8.0±3.0 | 317 | 259 (81.7) | 94.0 | 0.5 | 0.91 | <0.001 |
| Tricuspid regurgitation maximum velocity, m/s | 2.35±0.39 | 2.29±0.35 | 272 | 260 (95.6) | 94.9 | 0.3 | 0.83 | <0.001 |
| LAVI, mL/m2 | 26.6±11.3 | 26.6±11.3 | 289 | 246 (85.1) | 98.6 | 0.7 | >0.99 | <0.001 |
| Aortic valve maximum velocity, m/s | 1.42±0.39 | 1.41±0.38 | 317 | 313 (98.7) | 97.8 | 0.2 | 0.95 | <0.001 |
| Aortic valve mean pressure gradient, mm Hg | 5.0±3.3 | 4.8±3.2 | 317 | 313 (98.7) | 98.7 | 1.6 | 0.96 | <0.001 |
| Global longitudinal strain,‡ % | 19.7±3.4 | NA | 296 | 285 (96.3) | NA | NA | NA | NA |
| TAPSE, mm | 21.4±5.0 | 21.5±4.9 | 136 | 135 (99.3) | 98.5 | 0.7 | >0.99 | <0.001 |
AI indicates artificial intelligence; E/A, early to late ventricular filling velocity ratio; E/e’ (lateral), ratio of early diastolic velocity to lateral mitral annular velocity; E/e’ (septal), ratio of early diastolic velocity to septal mitral annular velocity; e’ (lateral), lateral mitral annular velocity; e’ (septal), septal mitral annular velocity; IVSTd, interventricular septal thickness in diastole; LAVI, left atrial volume index; LVEDV, left ventricular end‐diastolic volume; LVEF, left ventricular ejection fraction; LVESV, left ventricular end‐systolic volume; LVIDd, left ventricular internal diameter in diastole; LVIDs, left ventricular internal diameter in systole; MV‐A, mitral valve late diastolic velocity; MV‐E, mitral valve early diastolic velocity; PWTd, posterior wall thickness in diastole; and TAPSE, tricuspid annular plane systolic excursion.
Data were considered acceptable if the difference between AI’s initial measurements and the final report values were within acceptable ranges listed in Table S4.
The average absolute modification by the sonographer and echocardiologists during the reevaluation of AI’s initial analysis. This metric represents the magnitude of adjustments performed by the sonographer to ensure the accuracy of the final report.
Global longitudinal strain was not routinely included in clinical reports.
Once the images were analyzed by AI, sonographers and echocardiologists reviewed the parameters and adjusted them for the final report if necessary. Notably, AI’s initial measurements were deemed acceptable for clinical use without significant modification in >90% of cases for nearly all parameters, with the exception of LV internal diameter (in diastole, 86.7%; in systole, 70.6%; definitions of the acceptable ranges are summarized in Table S4). The mean absolute modification by sonographers, calculated as the sum of the absolute differences between AI’s initial measurements and the final report values divided by the number of modified AI values, showed that the modifications required were small, even when necessary. For example, the mean absolute modification was 3.9% for LV ejection fraction and 3.4% for LV end‐diastolic diameter.
Scatterplots and Bland–Altman plots were used to visualize the systematic errors between AI’s initial measurements and the final report values, as shown in Figure 3 and Figure S2. Intraclass correlation coefficients between AI and sonographers exceeded 0.8 for all parameters, as shown in Table 2, with P values <0.001, indicating a high level of agreement. For systolic and diastolic LV diameters and wall thicknesses, it was observed that AI tended to return more extreme values; that is, larger values were slightly overestimated and smaller values were slightly underestimated.
Figure 3. Concordance of AI’s initial measurements and the final values.
The scatter plots (left) demonstrate a strong correlation between AI‐generated data and the final report values. The Bland–Altman plots illustrate the mean differences (bias) and LOAs between the 2 measurement sets, with bias close to zero for each parameter. The overall concordance was excellent, with ICCs exceeding 0.8 and narrow LOAs. AI indicates artificial intelligence; E/A, early to late ventricular filling velocity ratio; ICC, intraclass correlation coefficient; LAVI, left atrial volume index; LOA, limits of agreement; LVEF, left ventricular ejection fraction; and LVIDd, left ventricular internal diameter in diastole.
Discussion
This first‐ever trial to randomize the use of an AI‐based automated tool on a daily basis revealed that AI significantly enhanced the efficiency of screening echocardiography, reducing examination time despite a 3.4‐fold increase in the number of parameters measured. This improved efficiency increased the number of examinations per day without increasing sonographers’ fatigue; in fact, it mitigated mental fatigue. Furthermore, being freed from the need to perform time‐consuming measurements allowed sonographers to focus on image acquisition, which led to improved image quality.
AI‐based tools in echocardiography have been extensively investigated since approximately 2018, 20 , 21 with numerous studies demonstrating their accuracy and efficiency in analyzing echocardiographic images. For example, a retrospective study using the same AI tool used in our trial confirmed its capacity to accurately analyze echocardiographic parameters under controlled conditions. 11 Another experimental study reported a 70% reduction in analysis and reporting time with this AI tool. 9 Despite these promising results, AI has not yet achieved widespread adoption in clinical practice. A significant barrier to broader implementation is that many studies evaluating AI tools, including those mentioned above, were conducted in retrospective or experimental settings, which do not fully capture the operational complexities of routine clinical workflows.
Recently, several studies have begun addressing this gap by evaluating AI tools in prospective, randomized trials. 22 , 23 In contrast with retrospective studies, which lack the complexity of real‐world clinical settings, these randomized trials provide more practical insights. One notable example is EchoNet‐RCT (Blinded, Randomized Controlled Trial of Sonographer Versus Artificial Intelligence Assessment of Cardiac Function), the first randomized controlled trial to evaluate AI in echocardiography, which demonstrated the precision of AI in automatically analyzing echocardiographic images to measure LV ejection fraction, and validated the accuracy of AI‐guided workflows in clinical practice. 8 Our study, using an AI tool capable of analyzing a broad range of clinical echocardiographic parameters, extends this focus beyond AI accuracy to examine its broader clinical impact.
Specifically, our findings demonstrated that integrating AI into daily echocardiography workflows significantly enhanced efficiency by reducing total examination time while simultaneously increasing the number of examinations performed per day. On AI days, sonographers were able to complete more examinations without sacrificing diagnostic quality. The AI tool not only automated the analysis of routine echocardiographic parameters but also facilitated the acquisition of more comprehensive data, including complex measurements like LV strain, which are not typically analyzed during routine screening. Importantly, the diagnostic integrity of these AI‐generated measurements was maintained, as evidenced by the high concordance between AI’s initial measurements and the final expert‐reviewed reports. This consistency underscores AI’s potential to augment sonographers’ capacity for routine echocardiographic assessments, allowing for increased throughput without compromising clinical accuracy.
A critical finding of our study was the potential for AI to alleviate the mental and physical burden on sonographers, as indicated by the questionnaire results—an especially relevant benefit in high‐volume echocardiographic laboratories. Screening echocardiography, often viewed as repetitive and routine, requires sonographers to conduct rapid yet accurate assessments. These repetitive tasks, combined with increasing clinical demands, may reduce motivation and increase fatigue among sonographers. 2 Our results suggest that AI can help mitigate these challenges. Moreover, while the time saved through AI‐enhanced efficiency in our study was used to perform more examinations, this additional time could be repurposed for more complex clinical activities, such as discussing hemodynamic findings or treatment strategies with clinicians, or providing more detailed explanations to patients. Engaging in these more intellectually stimulating and patient‐centered tasks could enhance job satisfaction for sonographers, contributing to a more rewarding and sustainable clinical practice.
Another significant outcome from our trial was the improvement in the quality of echocardiographic images on AI‐assisted days. High‐quality imaging is crucial for ensuring diagnostic accuracy, particularly for clinicians reviewing the images after acquisition. It is likely that the observed improvement in image quality on AI days occurred because sonographers were able to devote more attention to image acquisition when relieved of the cognitive load of performing manual measurements. Additionally, sonographers likely recognized that AI algorithms perform optimally with high‐quality images, motivating them to ensure the best possible image acquisition under AI guidance, even in cases where they might have previously accepted suboptimal images when manually handling measurements. This enhancement in image quality not only facilitates better AI performance but also ensures that subsequent clinical decisions are based on clearer, more accurate data, further underscoring the value of integrating AI into routine echocardiographic practice. However, because the improvement was mainly a shift from “good” to “excellent” rather than a reduction in poor‐quality images, the direct clinical impact of this finding on diagnostic accuracy and reproducibility remains to be determined.
Looking ahead, the integration of AI into echocardiographic practice presents a promising opportunity to enhance workflow efficiency, reduce sonographers’ burden, and potentially support more personalized and comprehensive patient care. Further refinement of AI tools and larger‐scale studies will be necessary to fully assess their clinical utility and pave the way for broader adoption in routine practice.
Future directions include conducting multicenter trials with longer follow‐up to confirm the sustainability and generalizability of our findings, evaluating the potential impact of AI among junior sonographers or trainees where the benefits may be even greater, and assessing patient‐centered outcomes in specific populations such as individuals with heart failure, in whom more frequent echocardiographic follow‐up could improve diagnostic accuracy, guide treatment changes, and ultimately influence prognosis.
Limitations
This study has several limitations. First, it was conducted at a single center, and since echocardiography protocols can differ across institutions, multicenter studies are needed to confirm the generalizability of these results. Additionally, the study duration was limited to about 2.5 months, so it remains unclear whether the observed efficiency gains, workload improvements, and enhancements in image quality would persist over a longer period. Second, although the trial was randomized, sonographers were aware of AI usage, which could have influenced their performance, even though they were blinded to the study’s end points. This awareness may have introduced some bias. Third, this study focused on experienced sonographers, leaving open the question of whether similar improvements in time efficiency and outcomes would be observed with less experienced operators. Prior research suggests that AI may enhance performance even for novices, 24 but fundamental sonography skills might still be required to fully leverage the benefits of AI. Next, our study focused on workflow efficiency and diagnostic quality but did not directly assess how AI‐assisted echocardiography impacts patient outcomes or clinical decision making. Furthermore, the Likert scale questionnaire that we used to measure sonographers’ fatigue was our original and had not been validated, and the reliance solely on subjective self‐reported scores may limit the robustness of the findings. Finally, the majority of patients included were undergoing screening echocardiography, primarily in sinus rhythm and with low rates of complex cardiac pathology. This may limit the applicability of our findings to more complex clinical cases, such as those with advanced heart failure or significant valvular disease, where human oversight might be more critical. Further research is needed to assess the impact of AI on both workflow efficiency and clinical outcomes in a broader range of clinical scenarios.
Conclusions
This randomized trial demonstrated that integrating an AI‐based automated analysis system into clinical echocardiography workflows can significantly reduce examination time and increase the number of examinations performed per day, thereby enhancing the overall efficiency of the echocardiography laboratory. Importantly, AI analysis maintained diagnostic quality, mitigated sonographers’ fatigue, and improved image quality, supporting its potential role in streamlining clinical practice. Continued advancements in AI technology hold promise for further improvements in both workflow efficiency and diagnostic accuracy in echocardiography.
Sources of Funding
This study was partially supported by the Japan Society for the Promotion of Science KAKENHI (grant number 25K19371). M3AI, the distributor of the Us2.ai platform in Japan, provided part of the software capabilities free of charge, and 2 personnel from M3AI participated in the study as coauthors. However, the study was audited and monitored by a researcher not involved in the study to ensure that the results were not influenced for the benefit of the company. Importantly, M3AI had no role in data collection, statistical analysis, or interpretation of the results, all of which were conducted independently by the academic investigators. No financial compensation was made to the company or the researchers involved.
Disclosures
Dr Kagiyama received research grants from AstraZeneca, EchoNous Inc., Bristio Myers Squibb, and AMI Inc.; and speaker honoraria from Eli Lilly, Novartis, Otsuka pharmaceutical, Bristol Myers Squibb, and Boehringer‐Ingelheim outside this work, and was affiliated with a department funded by Paramount Bed Ltd. Dr Kaneko received research grants from the Japanese Circulation Society and speaker honoraria from Abbott Medical Japan, and was affiliated with a department funded by Abeam Consulting. The remaining authors have no disclosures to report.
Supporting information
Tables S1–S4
Figures S1–S2
Acknowledgments
The authors thank the ethics committee of the Faculty of Medicine, Juntendo University, for their approval and oversight. We also acknowledge M3AI Inc. for providing technical support related to the AI analysis software and thank the clinical staff and patients who participated in this study. N.K. and T.M. supervised the study. N.K. designed the study. A.S., N.K., E.S., Y.N., A.M., T.K., and S.M. conducted the study. A.S. and N.K. performed the statistical analyses. A.S. and E.S. assessed image quality. Y.A. and K.S. coordinated the implementation of the AI analysis software. All authors had access to the data presented in the study, contributed to critical revisions, and approved the final version of the manuscript. All authors take responsibility for the integrity of the work and were responsible for the decision to submit for publication.
This study was presented at the at American Heart Association Scientific Sessions on November 16–18, 2024, in Chicago, IL.
This manuscript was sent to Daniel E. Clark, MD, MPH, Associate Editor, for review by expert referees, editorial decision, and final disposition.
Preprint posted on MedRxiv August 24, 2025. doi: https://doi.org/10.1101/2025.08.20.25334115.
Supplemental Material is available at https://www.ahajournals.org/doi/suppl/10.1161/JAHA.125.045637
For Sources of Funding and Disclosures, see page 10.
References
- 1. Papolos A, Narula J, Bavishi C, Chaudhry FA, Sengupta PP. U.S. hospital use of echocardiography. J Am Coll Cardiol. 2016;67:502–511. doi: 10.1016/j.jacc.2015.10.090 [DOI] [PubMed] [Google Scholar]
- 2. Younan K, Walkley D, Quinton AE, Alphonse J. Burnout in the sonographic environment: The identification and exploration of the causes of sonographer burnout and strategies for prevention and control. Sonography. 2022;9:175–185. doi: 10.1002/sono.12333 [DOI] [Google Scholar]
- 3. Kusunose K, Okushi Y, Okayama Y, Zheng R, Nakai M, Sumita Y, Ise T, Yamaguchi K, Yagi S, Fukuda D, et al. Use of echocardiography and heart failure in‐hospital mortality from registry data in Japan. J Cardiovasc Dev Dis. 2021;8:124. doi: 10.3390/jcdd8100124 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4. Kagiyama N, Shrestha S, Farjo PD, Sengupta PP. Artificial intelligence: practical primer for clinical research in cardiovascular disease. J Am Heart Assoc. 2019;8:e012788. doi: 10.1161/jaha.119.012788 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5. Tamura Y, Nomura A, Kagiyama N, Mizuno A, Node K. Digitalomics, digital intervention, and designing future: The next frontier in cardiology. J Cardiol. 2024;83:318–322. doi: 10.1016/j.jjcc.2023.12.002 [DOI] [PubMed] [Google Scholar]
- 6. Komuro J, Kusumoto D, Hashimoto H, Yuasa S. Machine learning in cardiology: clinical application and basic research. J Cardiol. 2023;82:128–133. doi: 10.1016/j.jjcc.2023.04.020 [DOI] [PubMed] [Google Scholar]
- 7. Myhre PL, Hung C‐L, Frost MJ, Jiang Z, Ouwerkerk W, Teramoto K, Svedlund S, Saraste A, Hage C, Tan R‐S, et al. External validation of a deep learning algorithm for automated echocardiographic strain measurements. Eur Heart J Digit Health. 2023;5:60–68. doi: 10.1093/ehjdh/ztad072 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8. He B, Kwan AC, Cho JH, Yuan N, Pollick C, Shiota T, Ebinger J, Bello NA, Wei J, Josan K, et al. Blinded, randomized trial of sonographer versus AI cardiac function assessment. Nature. 2023;616:520–524. doi: 10.1038/s41586-023-05947-3 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9. Hirata Y, Nomura Y, Saijo Y, Sata M, Kusunose K. Reducing echocardiographic examination time through routine use of fully automated software: a comparative study of measurement and report creation time. J Echocardiog. 2024;22:162–170. doi: 10.1007/s12574-023-00636-6 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10. Ouyang D, He B, Ghorbani A, Yuan N, Ebinger J, Langlotz CP, Heidenreich PA, Harrington RA, Liang DH, Ashley EA, et al. Video‐based AI for beat‐to‐beat assessment of cardiac function. Nature. 2020;580:252–256. doi: 10.1038/s41586-020-2145-8 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11. Tromp J, Bauer D, Claggett BL, Frost M, Iversen MB, Prasad N, Petrie MC, Larson MG, Ezekowitz JA, Solomon SD. A formal validation of a deep learning‐based automated workflow for the interpretation of the echocardiogram. Nat Commun. 2022;13:6776. doi: 10.1038/s41467-022-34245-1 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12. Tromp J, Seekings PJ, Hung CL, Iversen MB, Frost MJ, Ouwerkerk W, Jiang Z, Eisenhaber F, Goh RSM, Zhao H, et al. Automated interpretation of systolic and diastolic function on the echocardiogram: a multicohort study. Lancet Digit Health. 2022;4:e46–e54. doi: 10.1016/s2589-7500(21)00235-1 [DOI] [PubMed] [Google Scholar]
- 13. Topol EJ. High‐performance medicine: the convergence of human and artificial intelligence. Nat Med. 2019;25:44–56. doi: 10.1038/s41591-018-0300-7 [DOI] [PubMed] [Google Scholar]
- 14. Moher D, Schulz KF, Altman DG. The CONSORT statement: revised recommendations for improving the quality of reports of parallel‐group randomized trials. Ann Intern Med. 2001;134:657–662. doi: 10.7326/0003-4819-134-8-200104170-00011 [DOI] [PubMed] [Google Scholar]
- 15. Lang RM, Badano LP, Mor‐Avi V, Afilalo J, Armstrong A, Ernande L, Flachskampf FA, Foster E, Goldstein SA, Kuznetsova T, et al. Recommendations for cardiac chamber quantification by echocardiography in adults: an update from the American Society of Echocardiography and the European Association of Cardiovascular Imaging. J Am Soc Echocardiog: Offic Public Am Soc Echocardiog. 2015;28:1–39. doi: 10.1016/j.echo.2014.10.003 [DOI] [PubMed] [Google Scholar]
- 16. Mitchell C, Rahko PS, Blauwet LA, Canaday B, Finstuen JA, Foster MC, Horton K, Ogunyankin KO, Palma RA, Velazquez EJ. Guidelines for performing a comprehensive transthoracic echocardiographic examination in adults: recommendations from the American Society of Echocardiography. J Am Soc Echocardiog Offic Public Am Soc Echocardiog. 2019;32:1–64. doi: 10.1016/j.echo.2018.06.004 [DOI] [PubMed] [Google Scholar]
- 17. Nagata Y, Kado Y, Onoue T, Otani K, Nakazono A, Otsuji Y, Takeuchi M. Impact of image quality on reliability of the measurements of left ventricular systolic function and global longitudinal strain in 2D echocardiography. Echo Res Pract. 2018;5:27–39. doi: 10.1530/erp-17-0047 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18. Daimon M, Watanabe H, Abe Y, Hirata K, Hozumi T, Ishii K, Ito H, Iwakura K, Izumi C, Matsuzaki M, et al. Normal values of echocardiographic parameters in relation to age in a healthy Japanese population: the JAMP study. Circul J: Offic J Japan Circul Soc. 2008;72:1859–1866. doi: 10.1253/circj.cj-08-0171 [DOI] [PubMed] [Google Scholar]
- 19. Asch FM, Miyoshi T, Addetia K, Citro R, Daimon M, Desale S, Fajardo PG, Kasliwal RR, Kirkpatrick JN, Monaghan MJ, et al. Similarities and differences in left ventricular size and function among races and nationalities: results of the world Alliance societies of echocardiography Normal values Study. J Am Soc Echocardiog: Offic Public Am Soc Echocardiog. 2019;32:1396–1406.e1392. doi: 10.1016/j.echo.2019.08.012 [DOI] [PubMed] [Google Scholar]
- 20. Madani A, Arnaout R, Mofrad M, Arnaout R. Fast and accurate view classification of echocardiograms using deep learning. NPJ Digital Med. 2018;1:1. doi: 10.1038/s41746-017-0013-1 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21. Zhang J, Gajjala S, Agrawal P, Tison GH, Hallock LA, Beussink‐Nelson L, Lassen MH, Fan E, Aras MA, Jordan C, et al. Fully automated echocardiogram interpretation in clinical practice. Circulation. 2018;138:1623–1635. doi: 10.1161/circulationaha.118.034338 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22. Woodward G, Bajre M, Bhattacharyya S, Breen M, Chiocchia V, Dawes H, Dehbi HM, Descamps T, Frangou E, Fazakarley CA, et al. PROTEUS Study: a prospective randomized controlled trial evaluating the use of artificial intelligence in stress echocardiography. Am Heart J. 2023;263:123–132. doi: 10.1016/j.ahj.2023.05.003 [DOI] [PubMed] [Google Scholar]
- 23. Khan E, Lambrakis K, Briffa T, Cullen L, Karnon J, Papendick C, Quinn S, Tideman P, Van Den Hengel A, Verjans J, et al. Re‐engineering the clinical approach to suspected cardiac chest pain assessment in the emergency department by expediting research evidence to practice using artificial intelligence (RAPIDx AI)‐a cluster randomised clinical trial design. Heart Lung Circ. 2022;31:S224–S225. doi: 10.1016/j.hlc.2022.06.373 [DOI] [PubMed] [Google Scholar]
- 24. Narang A, Bae R, Hong H, Thomas Y, Surette S, Cadieu C, Chaudhry A, Martin RP, McCarthy PM, Rubenson DS, et al. Utility of a deep‐learning algorithm to guide novices to acquire echocardiograms for limited diagnostic use. JAMA Ccard. 2021;6:624–632. doi: 10.1001/jamacardio.2021.0185 [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Tables S1–S4
Figures S1–S2
Data Availability Statement
The data underlying this article cannot be shared publicly for the reason of maintaining the privacy of individuals who participated in the study. The data will be shared upon reasonable request to the corresponding author.