The effect of AI-assisted bedside echocardiography on inpatient care: a prospective trial

Abstract

Aims

Focused cardiac ultrasound (FoCUS) can yield valuable information for decision-making. However, it is limited by the skills required to acquire and interpret high-quality images. Machine learning algorithms can help mitigate this gap by providing guidance for optimal image acquisition and interpretation. We aimed to evaluate the impact of an artificial intelligence (AI) assisted, FDA-cleared FoCUS platform on clinical decision-making.

Methods and results

This was a prospective trial with pre and post sequential allocation, conducted in two internal medicine departments. During the first 2 months, physicians with no formal echocardiography training used a common commercial FoCUS device as a complementary tool for their bedside patient evaluations. Then, during the following 4 months, an AI cloud-based platform was added, providing real-time feedback for image acquisition and AI-based echocardiographic results. The primary outcome was change in care following FoCUS, as reported by physicians after the examination and verified by assessors, which was analysed by generalized linear mixed model accounting for physician and department effects. Two hundred and eighty-one patients met the inclusion criteria and underwent FoCUS, 110 (39%) without AI assistance (control) and 171 with the AI. The most common reasons for FoCUS were worsening dyspnoea (50%) and chest pain (20%). A non-significant trend was observed in physician-reported new echocardiographic findings towards the AI group (43% vs.34%, P = 0.11). The FoCUS led to a change of care more often in the AI group (32% vs. 20%, adjusted OR 1.87, 95% CI 1.05–3.32). The number needed to scan with the AI to have one additional change in care was 9 (95% CI 5–57). In multivariate analysis, AI use was an independent predictor for a FoCUS-led change of care (adjusted OR 2.16, 95% CI 1.18–3.97), an effect that consisted of subgroup analysis and an interrupted time-series model. AI also led to a lower rate of inpatient formal echocardiographic examinations (43% vs. 27%, P = 0.006)

Conclusion

AI-assisted FoCUS led to a higher rate of treatment plan changes, highlighting its potential to enhance bedside cardiac evaluation and optimize patient management.

Graphical Abstract

Graphical Abstract.

Introduction

Echocardiography is the cornerstone of cardiovascular disease diagnosis and management.¹ In recent years, this non-invasive test has become an invaluable tool for patient assessment in various clinical settings, offering insights regarding structural heart disease, hemodynamics, and aetiologies for the deteriorating patient.^2,3 The growing need for echocardiography as a decision-making tool, combined with technological advances in device size and cost, led to the widespread use of point-of-care focused cardiac ultrasound (FoCUS).^4,5 Clinicians can benefit from the use of such point-of-care devices in various clinical scenarios, especially where prompt evaluation of cardiac structure and function is needed, as is often the case in the inpatient setting.^6,7 However, the use of FoCUS is limited by the high level of expertise needed for image acquisition and interpretation, leading to potential delays in diagnosis and treatment.⁸

Recent technological advancements in the field of machine learning (ML) and artificial intelligence (AI) have made the application of AI modules to different imaging modalities, and specifically echocardiography, feasible.^9,10 These new tools have begun to transform the field of echocardiography with new AI-driven tools that offer image acquisition guidance, automated measurements, and decision support tools.^10–12 Moreover, AI assistance enables physicians with limited experience to accurately assess basic parameters such as left ventricular ejection fraction (LVEF) and even the assessment of diastolic dysfunction with comparable results to those of highly experienced peers.^13–16

While previously mentioned studies are promising, only a few evaluated the effects of AI assistance on ‘real-world’ clinical decision-making and most of them were focused on a single pathology or finding. Therefore, we aimed to explore the implications of incorporating an FDA-cleared AI-assisted FoCUS that provides simultaneous comprehensive cardiac evaluation into routine patient care and to determine its impact on inpatient management clinical decision-making.

Methods

Study population and design

This is a prospective study that took place in two internal medicine departments at a large tertiary referral hospital in Israel during 2024. The study included eight internal medicine physicians (four from each department) in different stages of their residency with no formal echocardiography training. The study was conducted according to the Declaration of Helsinki and approved by the Tel-Aviv Sourasky Medical Centre review board (TLV-0198-24). The study protocol was also submitted prior to initiation to the Ministry of Health national clinical trial database (MOH_2024-05-21_013425). Physicians signed an informed consent, while patients gave verbal consent, considering that the intervention was part of routine care delivered by the included physicians.

Participating physicians were residents at various stages of their training, where the residency programme does not require any level of competency in FoCUS examinations, nor does it provide formal training in this field. While some residents may have used POCUS as part of their routine clinical evaluations prior to the trial, bedside echocardiography in particular is rarely performed, apart from IVC assessment for evaluating congestion. This information was verified by the study’s recruiting team through interviews and a basic competency assessment conducted by an expert cardiologist. To establish an even baseline, all participating physicians received a 4-hour dedicated training targeting the operation of the ultrasound system (Lumify^™, Philips, Amsterdam, The Netherlands), the software (AISAP-POCAD^TM, Ramat Gan, Israel), and basic echocardiographic view acquisition, that were demonstrated by a trained echocardiography technician for standardization. This was followed by a training session on human models under the guidance of a cardiologist.

Patients who were under the care of one of the participating physicians were routinely and consecutively screened for this study. Patients included in this study were all admitted to internal medicine departments with one of the following symptoms or conditions as determined by the admitting IM physician in the department—worsening dyspnoea, chest pain, syncope, suspected heart failure, suspected arrhythmia, suspected valvular disease, or a different condition that required cardiac evaluation. Patients with prosthetic valves were excluded due to the technical difficulty of the chosen algorithm.

In order to mitigate any effect of the chosen device or software, other than the added value of AI support, we used a pre-post design as detailed below.

Study protocol

During the study period, the included physicians used FoCUS as part of their routine evaluation of patients who met the inclusion criteria as defined above. The FoCUS evaluation included 4 common views—parasternal long axis with and without Doppler, parasternal short axis (at the aortic valve level), apical four-chamber view with and without Doppler, and subcostal view. Physicians approached relevant patients and explained about the device, AI use (in the intervention), and general instructions for the examination. Once the patient gave oral consent, the physician performed FoCUS using the studied device alongside routine physical examination. Physicians were required to assess the following FoCUS findings—left ventricular ejection fraction (LVEF), valvular abnormalities (including severity), pericardial effusion, and IVC diameter and overall LV mass assessment.

The study was performed using a pre-post design (Figure 1). During the first 2 months (control), FoCUS was performed without AI-assistance. Therefore, physicians scanned independently and produced an exam report (see Supplementary material online, Figure S1). In the following 4 months (interventions), the users received real-time feedback for image acquisition using an on-screen quality score, and, at the end of the exam, an AI-generated exam summary was offered for the physician to review (a video portraying the scan and report process is available in the Supplementary material online). Of note, when using the FoCUS with AI assistance, a minimal quality image had to be obtained in each plane for the AI to provide the full final report (whether presented to the physician or not), although it was not an obligatory step for completion of FoCUS. In order to mitigate any effects of the chosen device or software, other than the added value of AI support, the exact same hardware and software were used throughout the study. Exam reports, evaluation dates, and scanning times were coded and saved on a secure cloud-based storage.

Figure 1.

Study design and timeline.

AI-based algorithm

The algorithm used in this study is FDA cleared (K234141) and developed by AISAP LTD (Ramat Gan, Israel). Extended information regarding module training and validation studies and performance can be found in our Supplementary material online. Briefly, the AISAP software integrates deep learning models for automated analysis of echocardiographic studies by analysing uploaded DICOM files. The first component is a view classification model based on the EfficientNetV2 architecture, trained to assign each video loop to one of the main TTE views. Accurate view classification enables downstream models to operate within the correct anatomical and clinical context.

The second component includes a set of specialized models designed for diagnosis and measurement tasks. These include (1) a diagnosis classification model combining CNN's and transformers, trained using cross-entropy loss to predict categorical disease severity (e.g. valvular disease, right ventricular dysfunction); (2) a regression model with similar architecture trained using L2 loss to estimate continuous clinical parameters such as left ventricular ejection fraction; and (3) an instance segmentation model that processes all frames of relevant views to segment anatomical structures, identify key frames, and apply logic-based rules to extract and validate measurements.

Study outcomes

To evaluate the effect of the FoCUS, a questionnaire was filled immediately after the scan (up to 120 min from scan completion) by the physician (see Supplementary material online, Table S1, supplement). Our primary outcome was a change in treatment plan that stemmed directly from the FoCUS scan. These changes included one or more of the mentioned below: initiation or modification of medications, recommendation for invasive procedure performed during or adjacent to admission (valve replacement or coronary angiography), completion of cardiac coronary single-photon emission computed tomography (SPECT) or coronary CT angiography before discharge, and changes in patient placement (transfer to cardiology or intensive care units). The primary outcome was validated at the end of the trial by independent assessors (O.F. and R.M., attending internal medicine physicians) who were masked to group allocation. For validation, medical files (follow-up notes, discharge summary, in-hospital medication lists) from patients’ hospitalization were printed and provided to the assessors without dates, with verification that no information regarding the allocation was mentioned in the files. The assessors were provided with the presumed change in care following the bedside echo for each patient. All cases were reviewed separately by each assessor, and in cases of disagreement (6 cases out of 76, 8%), files were reviewed together until a final decision was made.

Our secondary outcomes included the rate of physician-reported new findings following the FoCUS and the rate of formal echocardiography exams performed before hospital discharge. New FoCUS findings were any of the following: 1) new LVEF ≤40%, 2) New mitral, tricuspid or aortic regurgitation and aortic stenosis, graded moderate and above, 3) pericardial effusion, 4) hypertrophic cardiomyopathy, and 5) non-collapsed inferior vena cava (IVC). A finding was considered ‘new’ only if it was not present in any prior formal echocardiography of the patient or was not known in patients without prior available echocardiography results (based on prior clinic visits and diagnoses).

Of note, to simulate a ‘real-world’ scenario, FoCUS findings were extracted from physicians’ post-exam questionnaires and not based on the AI/formal echocardiography. Therefore, they were based on physicians’ interpretation of the FoCUS images gained during their assessment (with or without AI support, based on the study group).

Data analysis

Association between the study group and the primary outcome is presented by odds ratio and 95% CI, using generalized linear mixed model (GLMM) with random intercepts for department and physician. We compared proportions of cases with a change in care using risk ratios and risk differences with 95% confidence intervals (CIs), and the number needed to scan (NNS) was calculated as the reciprocal of the absolute risk difference. Considering the possible impact of learning effects and shifts in case-mix, we performed a sensitivity analysis using interrupted time-series (ITS) utilizing monthly proportions of change in care, with adjustment for autocorrelation. To further account for a possible learning curve, we assessed the GLMM in a sub-group consisted of the last 10 exams of each physician in the control period and the first 10 exams of each physician in the intervention period (or fewer than 10 in each period for physicians with a lower number of assessments).

The primary outcome was also compared between cases of AI-based FoCUS (AI group) and the control group using multivariate logistic regression analysis. Covariates included as confounders were pre-specified, selected according to clinical relevance and baseline imbalances, and included: age (continuous), sex, physician department, history of heart failure and ischaemic heart disease, and indication for FoCUS. Continuous variables were compared by Mann–Whitney U tests, and Categorical variables by Chi-square tests. All analyses were performed in SPSS version 29.0, with a two-sided P-value less than 0.05 considered statistically significant. Confidence interval–based sample size calculations were assessed in R (version 4.3) using the PropCIs package.

As this was a real-world before–and–after study, no a priori power calculation was performed. Instead, a precision-based justification was applied. Based on the observed proportions of change in care (32% vs. 20%), the absolute risk difference of 12% was estimated with a 95% confidence interval of 1.8–22.2%. Under similar proportions and equal allocation, a total sample of ∼290 patients would achieve comparable precision (±10%).

Results

During the study period, 281 patients completed a FoCUS assessment by one of the included physicians (Figure 1). 110 (39%) assessments were without AI (control group) and 171 (61%) with the AI (AI group). Patients’ characteristics are shown in Table 1. The cohort's median (IQR) age was 75 (64–83) years, and 46% were females. The main reasons for FoCUS were worsening dyspnoea (50%), chest pain (20%), and syncope (11%), with no significant changes between the study groups (P = 0.106).

Table 1.

Cohort characteristics and comparison between the study groups

Variable	Total n = 281 (%)	Control n = 110 (%)	AI n = 171 (%)	P
Age	75 (64–83)	75 (61–82)	75 (65–84)	0.325
Female	129 (46)	38 (35)	91 (53)	0.006
Hyperlipidemia	162 (58)	69 (63)	93 (54)	0.167
Hypertension	179 (64)	71 (65)	108 (63)	0.813
Diabetes	110 (39)	48 (44)	62 (36)	0.216
Past/current smoker	77 (27)	29 (26)	48 (28)	0.780
COPD	33 (12)	13 (12)	20 (12)	0.975
Stroke	44 (16)	17 (16)	27 (16)	0.940
Heart failure	79 (28)	33 (30)	46 (27)	0.573
Ischaemic heart disease	95 (34)	38 (35)	47 (28)	0.208
Structural heart disease	72 (26)	30 (27)	42 (25)	0.611
Atrial fibrillation	61 (22)	21 (19)	40 (23)	0.393
Pacemaker	14 (5)	5 (5)	9 (5)	0.787
ICD	7 (3)	4 (4)	3 (2)	0.344
Reason for FoCUS				0.106
Worsening dyspnoea	139 (50)	53 (49)	86 (50)
Chest pain	56 (20)	18 (17)	41 (24)
Syncope	30 (11)	15 (14)	12 (7)
Susp. valvular disease	20 (7)	8 (7)	12 (7)
Others	35 (13)	15 (14)	20 (12)

The AI group acquired minimal-quality images for an AI report in 150 cases (88%), while the control obtained such quality images in 64 cases (58%, P < 0.001). The median (IQR) time of FoCUS scan was 5.9 min (4.4–7.7) in the AI group and 5.4 (2.6–7.8) in the control, without a statistical difference between the groups (P = 0.057).

Physicians reported new findings in 111 cases (40%). The findings included new reduced LVEF (45%), new moderate/severe valvular disorder (32%), distended IVC (21%), and two cases of hypertrophic cardiomyopathy (HCM). The rate of physician-reported new findings in the AI group was 43% and in the control group 34% (Figure 2), without a significant difference (P = 0.107). The type of new findings was similar between the groups (see Supplementary material online, Figure S2, P = 0.572).

Figure 2.

Sankey diagram of the study outcomes between the study groups.

FoCUS effect on patient care

FoCUS results have led to a change in care in 76 (28%) cases, including 57 (20%) pharmacological modifications and 22 (8%) non-pharmacological interventions (3 cases had both pharmacological and non-pharmacological changes). Non-pharmacological interventions included coronary catheterizations (n = 10), valvular interventions (n = 4), transfer to a higher level of care (cardiac or general intensive care unit, n = 4), and completion of cardiac SPECT (n = 4). The effect on care in each group is shown in Figure 3. The FoCUS results have more frequently led to a change in care in the AI group compared to the control (32% vs. 20%, adjusted OR 1.87, 95% CI 1.05–3.32). The corresponding risk ratio was 1.58 (95% CI 1.02–2.44), absolute risk difference 12% (95% CI 1.8–22%), and NNS with the AI to have one additional change in care was 9 (95% CI 5–57). The AI group had higher rates of both pharmacological and non-pharmacological interventions (23% vs. 16% and 10% vs. 5% accordingly), although these were not statistically significant.

Figure 3.

Percentage (95%CI) of changes in care observed between the study groups. Panel A—in the entire cohort; Panel B—subgroup of patients with newly reported echocardiographic findings.

Based on the ITS results, after the AI tool was introduced, there was an increase of about 17.5% points (β = 0.175, P = 0.032) in the proportion of cases with a change in care. The slope of change over time after AI (time slope = −0.024, P = 0.51) was not significantly different—meaning the effect was immediate rather than gradually increasing or decreasing. In the sub-group GLMM analysis among the 10 last control and 10 first intervention assessments for each physician, the AI continued to be associated with a change in care (adjusted OR 2.93, 95% CI 1.33–6.45, P = 0.008). In multivariate analysis (Table 2), the use of AI assistance was an independent predictor for a change in care (adjusted OR 2.16, 95% CI 1.18–3.97).

Table 2.

Univariate and multivariate analysis of predictors for change in care following FoCUS exam

Variable	Univariate		Multivariate
	OR (95% CI)	P	adjusted OR (95% CI)	P
AI group	1.85 (1.05–3.26)	0.03	2.16 (1.18–3.97)	0.13
Age, for every year	1.00 (0.98–1.02)	0.98	0.99 (0.98–1.02)	0.66
Female sex	1.01 (0.59–1.71)	0.98	0.90 (0.51–1.59)	0.71
Ischaemic heart disease	1.52 (0.88–2.62)	0.13	1.76 (0.92–3.37)	0.09
Heart failure	1.26 (0.71–2.36)	0.43	0.92 (0.47–1.82)	0.81
Department	1.50 (0.88–2.55)	0.13	1.52 (0.87–2.67)	0.15
Indication for FoCUS
Worsening dyspnoea	Ref.		Ref.
Chest pain	0.40 (0.19–0.86)	0.02	0.38 (0.17–0.83)	0.02
Syncope	0.41 (0.15–1.14)	0.09	0.40 (0.15–1.25)	0.12
Sus. valvular disease	0.63 (0.22–1.82)	0.39	0.57 (0.19–1.70)	0.31
Other	0.47 (0.19–1.15)	0.10	0.56 (0.22–1.41)	0.22

Following FoCUS, 33% of the patients were referred for inpatient formal echocardiography prior to discharge, with higher rates in the control group compared to the AI group (43% vs. 27%, P = 0.006). Cardiologic consultation was sought in 44 cases (16%), with similar rates between the control and AI groups (17% vs. 13%, P = 0.307).

Subgroup analysis

Considering that FoCUS examination without findings could also affect care (e.g. discontinuation of treatment), but might not be related to the AI use, we performed a subgroup analysis of patients with new physician-reported findings (n = 111). Patients in the AI group with FoCUS findings had higher odds for a change in care compared to the control (60% vs. 38%, OR 2.41, 95% CI 1.07–5.42, Figure 3B). In multivariate analysis (see Supplementary material online, Table S2), the use of AI remained an independent predictor for a change in care in this subgroup (aOR 3.15, 95% CI 1.22–8.13). In addition, patients with FoCUS findings in the AI group were less likely to perform a formal echocardiogram during hospitalization (31% vs. 51%, P = 0.038).

Discussion

This prospective study aimed to evaluate the effect of AI-assisted FoCUS on patient management and the decision-making process in real-word setting of internal medicine departments. Compared to a 2-month control period, the use of an AI-assisted FoCUS led to higher rates of patient management changes. This effect intensified in the subgroup of patients with positive FoCUS findings (60% vs. 38%) and was mainly driven by pharmacological changes made as a result of newly detected pathologies. Multivariate regression model revealed that AI-assistance was an independent predictor for performing a change in care, while it was associated with a reduced number of inpatient formal echocardiography examinations.

The use of AI-assistance in echocardiography is considered promising, primarily in the context of improved clinical workflows. Previous reports demonstrated that AI modules can reliably estimate basic echocardiographic parameters.^17–19 Furthermore, some studies suggest that AI and Machine learning ML based tools can mitigate the gap between highly experienced operators and relatively less experienced users regarding routine echocardiographic measurements.^19,20 The current study found that the use of AI-assisted FoCUS by internal medicine physicians led to a similar rate of new physician-reported findings (a non-statistically significant trend was observed), while simultaneously increasing patient management modifications. Thus, we further explore the concept that selected features in echocardiography can be performed by non-cardiologists given AI-assistance, especially when evaluating patients with symptoms that are suggestive of cardiac aetiology similar to patients chosen in this cohort.

There are few studies addressing the diagnostic yield of bedside cardiac ultrasound, especially those that include the use of AI modules in settings outside an echo laboratory.^21,22 There is also high variability in the definition of ‘new findings’ between different reports. However, studies addressing FoCUS as a screening tool with similar definitions to ours reported a detection rate of new findings of around 27%.²³ These reports are in concordance with our control results (34%), while the use of AI-assistance increased the rate of new findings to 43%. This trend might be the result of improvements in image acquisition enabled by AI acquisition quality, and to a higher degree of confidence in reporting a new pathology by the physician when there is an agreement with the AI module. This reassurance provided by AI modules was previously described in other fields where the use of AI bridged gaps and operated as a decision support tool.^24,25 It is important to note that the lack of significant differences between detection rates in the control and AI groups could stem from the fact that our study considered findings as new only if the operating physician reported them in the exam summary. Therefore, new FoCUS findings in the control group were those perceived by the physician based on the indexed examination. This simulated a real-world setting in which the physician makes decisions based on self-perceived new findings, without external validation, which was the main assessment of our study.

AI-related investigations have naturally targeted the accuracy and reliability of AI modules rather than their pure clinical effect. Studies addressing diagnostic yield and utility of formal echocardiography report clinical impact in under 30% of patients.²⁶ Information regarding other bedside modalities, such as general POCUS or lung ultrasound, resulted in a slightly higher number, ranging from 35–50%.^27,28 These differences may be attributed to a higher skill necessary for image acquisition or interpretation in echocardiography. Evidence regarding the integration of AI-assisted FoCUS by non-cardiologists in the context of clinical decision making is scarce. Therefore, our primary focus was to assess the impact of AI-assistance on patient care. It is interesting to evaluate the subgroup of patients with new findings, where 60% of scans with AI-assistance resulted in management modifications, while a significantly lower rate (38%) was found in the control group. These findings suggest that the value of AI assistance may reside not only in pure detection of pathologies but also in enhancing operator confidence, thus empowering clinicians. It is important to note that, unlike most previously published data, we performed a prospective design where video loops were reviewed, and decisions were made immediately after the FoCUS exam, strengthening the association between AI-assistance and treatment plan changes. Another interesting metric we were able to calculate was the number needed to scan (NNS) to produce one change in patient management. We found the NNS to be 9 (95% CI, 5–57). This specific parameter has been rarely reported in previous studies and, therefore, should be interpreted with caution, as our cohort was relatively small and consisted of a highly selected patient population with a predisposition to cardiac-related aetiologies.

Another interesting aspect of our study was the reduction in referrals for in-hospital formal echocardiography among patients in the AI-assisted FoCUS group (27% vs. 43%). This reduction provides a necessary reassurance for the concern that augmented use of bedside echocardiography will lead to an increased demand for an already limited resource. In this context, it is also worth mentioning that FoCUS scans performed with no AI assistance resulted in relatively low image quality as reflected by AI interpretation availability or cardiologist read-over, further enhancing our assumption that image acquisition and quality improvement contributed to physician confidence and guided treatment plan changes.

When conducting a study over a relatively long duration, the potential effect of a learning curve must be addressed. AI assistance can, and ideally should, contribute to a significant improvement in user proficiency over time. To account for this potential confounder and to validate the robustness of AI as an independent driver of treatment plan modifications, we performed a sub-analysis for each physician by comparing their last 10 scans without AI support to their first 10 scans with AI assistance. This analysis revealed a strong net effect of AI on clinical decision-making. However, we did not observe a statistically significant difference in scan duration between examinations performed with or without AI assistance. We hypothesize that the slightly longer scan time observed in the intervention group reflects the system’s requirement to achieve a minimum image quality threshold before generating a report.

Our study bears some notable limitations. First, being a single-centre study with only eight internal medicine physicians limits the generalisability of our results. Second, while our design was aimed at mitigating possible confounding factors of any new hardware or software, as well as past experience, the effect of possible improved skills over time cannot be excluded. While we included ITS and subgroup analyses, this should still be considered when interpreting our results. Third, physicians were not blinded to the use of AI, and the resulting bias cannot be ruled out. Fourth, our study population comprised inpatients with symptoms suggestive of a cardiac aetiology and therefore had a relatively high pretest probability. It is worth mentioning in the context that participating physicians were approximately 50–60% from each department's personnel, although patients are allocated randomly to any physician, some bias cannot be completely excluded. Fifth, the absence of a pre-specified power calculation is a limitation, and although our sample size achieved an approximate 95%CI of ±10% around the observed 12% absolute difference, future confirmatory studies designed for greater precision (e.g. ±5% CI) are needed. Finally, the effect on patient outcomes, including re-admission rate and mortality, was beyond the study scope and should be evaluated by future larger randomized controlled trials that account for baseline patient characteristics and indication for FoCUS.

In conclusion, this prospective controlled trial demonstrates that the use of AI-assisted FoCUS significantly influenced patient management, as reflected by higher rates of treatment modifications. These findings suggest improvements in imaging quality and operator confidence that can be attributed to AI performance as a decision support tool.

Supplementary material

Supplementary material is available at European Heart Journal – Digital Health.

Author’s contribution

SF- writing- original draft, data curation; OF- statistical analysis, writing—original draft; NB—writing—review and editing; NS—Data curation; RM –Project administration, Data curation; GB—Data curation; RM- writing—review and editing; UR- statistical analysis; OS-; MO- Project administration, validation; YT—writing—review and editing; RH- methodology, writing—review and editing; GB—conceptualisation, methodology, writing—review and editing.

Funding

Point-of-care ultrasound devices, as well as the cloud-based AI platform, were provided by AISAP LTD (www.aisap.ai).

Data availability

The data that support the findings of this study are not publicly available due to containing information that could compromise the privacy of research participants, but are available from the corresponding author, SF, upon reasonable request.

Ethics

The study protocol was reviewed and approved by the institutional ethics committee, approval number—TLV-24-0198.

Consent to participate statement

Written informed consent was obtained by writing from all participants. The study was conducted in accordance with the standards as laid down in the 1964 Declaration of Helsinki and its later amendments. Approved by the Tel-Aviv Sourasky medical center review board (TLV-24-198).