Deep learning-based echocardiography for cardiac assessment: a systematic literature review and meta-analysis

Abstract

Aims

Artificial intelligence, particularly deep learning (DL), has shown significant potential for enhancing the assessment of echocardiograms in clinical practice. This systematic review and meta-analysis aimed to evaluate the performance of DL models and to identify factors affecting their performance.

Methods and results

A systematic review of the PubMed, CENTRAL, and Web of Science databases was conducted, yielding 16 248 potential publications. Of these, 160 studies met the inclusion criteria. Only studies applying fully automated DL networks to echocardiographic loops were considered. Performance estimates were pooled using a random-effects meta-analysis. The studies were categorized into five distinct groups: prediction of disease or mortality, replication of established measures, segmentation of cardiac chambers, classification of chamber views, and assessment of image quality. High heterogeneity that could not be resolved was observed in studies focusing on cardiac chamber segmentation and classification. Model performance was influenced by a variety of different factors, including medical target and input characteristics such as video resolution, number of frames used, and total participants included during training.

Conclusion

DL models achieve high performance across various applications in echocardiography. In selected tasks, they may reduce reliance on expert input and support greater automation of echocardiographic workflows. However, further validation, assessment of clinical utility, and evaluation in real-world settings are required before widespread clinical adoption.

Graphical Abstract

Graphical Abstract.

Introduction

Echocardiography is a widely accessible, cost-effective, and non-invasive imaging technique that plays a pivotal role in assessing cardiac health.¹ It is routinely available in most clinical practices and offers real-time visualization of cardiac structure and function, making it indispensable in the diagnosis and management of numerous cardiovascular conditions.^2–4 Despite its extensive utility, the full diagnostic potential of echocardiography may remain untapped: while certain complex patterns and features might elude human interpretation, deep learning (DL) models are able to analyse these patterns⁵, representing an opportunity for advanced computational techniques to enhance its clinical value.^6,7

In recent years, DL models have demonstrated remarkable capabilities in analysing echocardiographic videos, detecting patterns, and deriving measurements that traditionally required human expertise.⁸ Furthermore, by leveraging the rich data embedded in ultrasound videos, these models can also extract novel metrics of cardiac health imperceptible to the human eye.⁹ This ability positions DL models as potential tools for providing rapid, standardized assessments of echocardiographic studies.¹⁰ Moreover, they hold transformative potential for advancing the utility of this technique by enabling the discovery of new insights into patient health while enhancing diagnostic precision.¹¹

In order to determine the potential of DL models as fully automated tools that expedite tasks traditionally done by physicians, a systematic literature review was conducted on the use of DL to assess cardiac health through the use of echocardiography alongside a meta-analysis of all comparable applications. Although numerous primary studies have reported promising results, the existing literature remains highly heterogeneous with respect to datasets, validation strategies, reported performance metrics, and even the types of tasks being assessed with DL. This heterogeneity makes it difficult to objectively interpret reproducibility and generalizability across clinical settings.

To date, published works have not performed a quantitative synthesis of performance of DL models on the same task across studies. A meta-analysis is therefore warranted to provide pooled performance estimates, assess between-study variability, and explore factors that may affect performance. Importantly, from a clinical perspective, individual high-performing models do not necessarily translate into reliable, scalable tools for real-world implementation. Quantitative synthesis therefore allows for a more rigorous evaluation of whether reported performance is consistent enough to support broader clinical adoption.

To our knowledge, no systematic review on DL in echocardiography, emphasizing what factors affect their performance, has yet been conducted. By pooling comparable outcomes, this study aims to provide a more objective and clinically meaningful assessment of the current maturity of fully automated DL-based echocardiographic tools, and to identify gaps where DL could feasibly be applied.

Methods

Systematic literature search

A systematic search was conducted using the online databases ‘PubMed’ (https://pubmed.ncbi.nlm.nih.gov/), ‘Cochrane Central Register of Controlled Trials’ (CENTRAL, https://www.cochranelibrary.com/central), and Web of Science (https://www.webofscience.com/). The search was limited to studies published after 2014 up to June 9th of 2026 to capture only novel works on DL. The search query used can be found in the Supplementary material online, Appendix  I, Methods.

Study selection

Eligibility criteria were developed collaboratively with a team of epidemiologists, cardiologists, and DL experts. Articles were considered eligible if they employed DL methods to analyse echocardiographic video data in a fully automated pipeline, defined as execution without manual pre-processing, user-guided feature selection, or expert interpretation at inference time. Given the large volume of DL research in echocardiography, inclusion was further restricted to studies in which the DL application could reasonably be considered a physician assistance tool, meaning that the outputs of the model directly supported clinical interpretation or decision making rather than image formation. Accordingly, studies focused exclusively on image enhancement or reconstruction tasks, such as denoising, dehazing, super-resolution, or image reconstruction, were excluded.

Articles were excluded if they met any of the following criteria: irrelevant content, no DL methods applied, absence of echocardiography videos as input, use of proprietary software, no reported performance against a ground-truth reference, or publication as a review, case report, letter, or other non-original research article. Studies relying on manual input prior to application of the DL model were excluded, including manual selection of video frames, manual extraction of still images from videos, reported manual view selection, manual region-of-interest selection used as model input, or any other expert-driven processing required prior to model application. Studies combining automated DL analysis with manual pre-processing were therefore not considered fully automated and were excluded. Manual annotations used solely for generation of ground-truth labels (e.g. expert defined regions-of-interest or segmentation) or training of the DL models were permitted. Studies analysing specific echocardiographic views were included provided that manual view assignment prior to model application was not explicitly reported. When the method of view assignment was not described, studies were retained to avoid excluding potentially fully automated approaches. Transoesophageal, intracardiac, foetal, paediatric, and animal echocardiography studies were excluded. Point-of-care ultrasound and stress echocardiography studies were eligible for inclusion.

Eligible studies were required to operate directly on echocardiogram videos or automatically extracted video-derived frames and to report a clearly defined analytical task. Tasks of interest included the automation of functions traditionally performed by experts, specifically: (i) segmentation of cardiac chambers; (ii) estimation or prediction of established quantitative echocardiographic measurements; (iii) cardiac chamber view classification; and (iv) assessment of image quality. In addition, studies applying DL models to diagnose disease or to predict clinical outcomes, including mortality, from echocardiographic data were included.

Titles and abstracts of all identified articles were screened for eligibility by two researchers (A.R.M. and S.Z.) using the online tool EPPI Reviewer (https://eppi.ioe.ac.uk/EPPIReviewer-Web/). Any uncertainties regarding an article’s eligibility were resolved in a structured consultation process with two additional reviewers—specialists in cardiology (S.-O.T. and P.S.W.). Full-text articles were assessed only if the title or abstract suggested relevance to the research topic. Reasons for considering studies ineligible for inclusion were recorded for both abstract and full article screening. Risk of bias was assessed using a modified version of the Risk of Bias Assessment Tool for non-randomized Studies.¹² Details are available in Supplementary material online, Appendix  I.

Data extraction

Results reported in this review were based on testing with external datasets whenever possible. If external testing was not done, results from the held-out data were recorded instead. In cases where comparable results from the same study were reported for held-out datasets but not for external datasets, the held-out testing results were included. Collected data were aggregated by DL goals (prediction of death or disease, established measure replication, segmentation, cardiac view classification, and quality assessment) to facilitate further analysis. The type of data extracted from all articles is available in the Supplementary material online, Appendix  I.

Statistical analysis

Sample sizes reported in the text refer explicitly to the number of individuals included in each study. In order to ensure consistent weighting across studies, those without information on the number of individuals in the testing set were excluded from the meta-analysis. Variance was calculated only for proportion-based metrics, using the reported performance estimate and the corresponding testing sample size. Proportions were transformed using the double-arcsine (Freeman-Tukey) method prior to meta-analysis to stabilize the variances of results reported in proportions. To ensure methodological consistency, meta-analyses were conducted separately for each performance metric, with studies pooled only when reporting the same measure (e.g. accuracy, area under the curve (AUC), Dice coefficient, mean absolute error, etc.) for comparable clinical tasks or outcomes. For non-proportion metrics for which standard errors, confidence intervals, or sufficient information to derive valid variance estimates were unavailable, inverse-variance pooling was not performed; these outcomes were summarized descriptively or analysed using the sample-weighted approaches specified in Supplementary material online, Appendix  I. Different metric types were not combined unless a statistically valid transformation enabled direct comparison. Information on specific transformations used to create comparable metrics can be found in the Supplementary material online, Appendix  I.

Variance due to heterogeneity between studies was reported using the I² statistic. Multilevel regression analyses were performed to address dependent data, in the case of this review, this occurred when a single study had multiple targets for their DL model, all with different effect sizes. In cases of residual heterogeneity, meta-regression analyses were performed if at least five studies reported relevant variables to test for moderation of results by characteristics describing the base architecture of the model, the size of the sample used for model training, the specific subgroup target, or any characteristic referring to the shape and size of the input. If sample size was not sufficient for analysis, moderators were applied independently. Cardiac chamber classification studies did not apply multilevel analysis due to independence of included works. Sensitivity analyses and outlier assessments were done using Cook’s distance (Cook’s D)¹³ to determine the undue influence of individual observations on model outcomes. Small-study effects were evaluated using funnel plots.¹⁴

Meta-analyses and meta-regression models were implemented using the R package metafor.¹⁵ A two-sided P-value below 0.05 was considered statistically significant. All statistical analyses were performed using the R statistical programming environment version 4.2.1. (https://www.r-project.org/).

Results

Study inclusion

A total of 16 248 studies were identified through abstract screening. After removing duplicates (n = 7903), 7710 articles were excluded based on title and abstract for not meeting inclusion criteria. A total of 635 full-text articles were assessed for eligibility. Ultimately, 160 articles, all of which can be found in the Supplementary material online, Appendix  I, were included in the final systematic review and meta-analysis (see Figure 1). Independent screening of a random sample of 103 abstracts by the two reviewers demonstrated 92.2% agreement (Cohen’s κ = 0.78, P < 0.001). The summarized characteristics of all included studies are presented in Supplementary material online, Table S1. Counts for DL model architectures stratified by task are presented in Table 1. U-Net was the most commonly used architecture for segmentation, whereas ResNet and three-dimensional convolutional neural networks (3D CNNs) were most frequently employed for disease or mortality prediction and the replication of established echocardiographic measurements. Clinical targets assessed within the three major task categories (disease or mortality prediction, replication of established measurements, and segmentation) are summarized in Table 2. Valvular heart diseases were the most frequently investigated target in prediction studies, while systolic function was the most commonly assessed target in studies aiming to replicate established echocardiographic measurements, most often being left ventricular ejection fraction (LVEF). Risk of bias assessments are provided in Supplementary material online, Figure S1, which show that studies from 2024 onward were less likely to be rated as having high risk of bias compared with earlier publications. A timeline depicting the use of each architecture from 2018 to 2026, along with key DL advancements during this period, is presented in Figure 2.^16–25

Figure 1.

Flowchart for study selection following PRISMA guidelines. DL, deep learning; PRISMA, preferred reporting items for systematic reviews and meta-analyses. The flowchart depicts the flow of information throughout each step of the systematic review.

Table 1.

Included studies aggregated by deep learning method used in each of the five tasks

	Total counts (N = 161)	Disease/mortality prediction (N = 59)	Measurement replication (N = 59)	Segmentation (N = 59)	Classification of views (N = 22)	Quality assessment (N = 10)
U-Net and variations	23.6 (38)	3.39 (2)	13.6 (8)	47.5 (28)	N/A	N/A
ResNet and variations	20.5 (33)	20.3 (12)	18.6 (11)	8.47 (5)	13.6 (3)	20.0 (2)
3D CNN	16.1 (26)	16.9 (10)	16.9 (10)	5.08 (3)	13.6 (3)	N/A
R2 + 1D	13.0 (21)	18.6 (11)	5.08 (3)	1.69 (1)	18.2 (4)	20.0 (2)
2D CNN	7.45 (12)	5.08 (3)	5.08 (3)	3.39 (2)	13.6 (3)	10.0 (1)
DeepLabV3	6.83 (11)	N/A	11.9 (7)	6.78 (4)	N/A	N/A
ViT	4.97 (8)	3.39 (2)	1.69 (1)	6.78 (4)	4.55 (1)	N/A
VGG	4.35 (7)	5.08 (3)	5.08 (3)	N/A	4.55 (1)	N/A
DenseNet	4.35 (7)	1.69 (1)	3.39 (2)	N/A	9.09 (2)	20.0 (2)
MobileNet	3.73 (6)	N/A	N/A	1.69 (1)	9.09 (2)	30.0 (3)
XceptionNet	2.48 (4)	N/A	1.69 (1)	3.39 (2)	4.55 (1)	N/A
PWC-Net	2.48 (4)	N/A	6.78 (4)	N/A	N/A	N/A
ConvNeXt	1.86 (3)	3.39 (2)	1.69 (1)	N/A	N/A	N/A
ViViT	1.86 (3)	3.39 (2)	1.69 (1)	N/A	N/A	N/A
Single use	14.9 (24)	13.6 (8)	8.47 (5)	15.3 (9)	9.09 (2)	N/A

Categories may overlap for individual studies. Categorical variables are presented as their relative (%) and absolute frequencies (n).

CNN, convolutional neural network; ViT, vision transformer; ViViT, video vision transformer; N/A, not applicable.

Table 2.

Included deep learning studies aggregated by clinical task being investigated, whether external validation was applied, and what type of training data were used

Clinical target per task	Counts % (n)
Prediction of disease or mortality	36.6 (59)
Valvular heart disease	35.6 (21)
Non-ischaemic cardiomyopathy	22.0 (13)
Other structural heart disease	16.9 (10)
Heart failure	16.9 (10)
Ischaemic cardiomyopathy	13.6 (8)
Pulmonary vascular disease	5.08 (3)
Medical device	5.08 (3)
Mortality	3.39 (2)
Atherosclerotic disease	1.69 (1)
Arrhythmias	1.69 (1)
Replication of established measures	36.6 (59)
Systolic function	71.2 (42)
Chamber volumes	18.6 (11)
Cardiac geometry	18.6 (11)
Event timing	15.3 (9)
Diastolic function	8.47 (5)
Demographic and anthropometrics	5.08 (3)
Circulating biomarkers	1.69 (1)
Segmentation	36.6 (59)
Cardiac chambers	89.8 (53)
Valves	15.3 (9)
Vessels	5.08 (3)
Validation type
External	53.4 (86)
Internal only	46.6 (75)
Reported sample size
Individuals	83.2 (134)
Echocardiographic videos	54.7 (88)
Echocardiographic studies	21.1 (34)

Categories may overlap for individual studies. Categorical variables are presented as their relative (%) and absolute frequencies (n). External datasets refer to independent datasets used for model testing only; studies in which the model was retrained or fine-tuned on the same dataset were not considered to have external validation.

Figure 2.

Timeline of published articles in the application of deep learning on echocardiography and their relation to advancements within the deep learning field. BERT, bidirectional encoder representations from transformers; ViT, vision transformer; CLIP, contrastive language-image pre-training; MAE, masked auto-encoder; SAM, segment anything. Classic CNN variants included 2D CNN, VGG, ResNet, DenseNet, EfficientNet, InceptionV3, MobileNetV2, XceptionNet, DCNN, 1D CNN, ConvNeXt, PWC-Net, and YoloV3, all of which are built on variations of convolutional layers for feature extraction in image or sequential data. Encoder-Decoders included U-Net, DeepLabV3, HRNet, Few-Shot Segmentation, and generic encoder modules, these networks are often used in segmentation and dense prediction. Spatiotemporal CNNs included 3D CNN, R2 + 1D, S3D, and SlowFast, all of which are designed to capture both spatial and temporal features for analysis of video or 3D data. Transformers included both ViT and ViViT (video vision transformer).

Availability of imaging views and clinically relevant information

Most studies focused on the apical 4-chamber view, with apical 2-chamber being the second most assessed view, followed closely by parasternal long axis (Figure 3). The total number of the cardiac views used in each study based on categorization of the model goal is shown in Supplementary material online, Figure S2. For those studies aiming to predict presence of disease or all-cause mortality over time, study participant characteristics were reported in 40 out of 59 articles. Further information on the proportion of reported information of interest is detailed in Supplementary material online, Table S2.

Figure 3.

Proportion of the usage of individual cardiac views per study. The x-axis indicates the proportion of studies that used each specific chamber view. The number inside of each bar is the exact number of studies that used that view for each separate task. A) Prediction of disease or mortality. B) Replication of established echocardiographic measures. C) Segmentation of cardiac chambers. D) Classification of cardiac chamber views. E) Assessment of echocardiogram image quality.

Prediction of prevalent disease or all-cause mortality

A multilevel meta-regression analysis of proportions for studies predicting prevalent disease or all-cause mortality was conducted in two groups: one reporting model performance as overall accuracy (n_studies = 22, n_targets = 36) and the other using AUC (n_studies = 23, n_targets = 70). Studies aggregated by target being predicted and ordered by accuracy are provided in Table 3. Residual heterogeneity decreased stepwise with the addition of moderators to the model. The inclusion of specific DL architectures resulted in high heterogeneity (I²_accuracy = 99.9%, I²_AUC = 98.8%), which decreased with the addition of the target (I²_accuracy = 98.4%, I²_AUC = 98.1%), and further decreased with the addition of participants used in model training, video resolution, number of frames, and batch size (I²_accuracy = 76.8%, I²_AUC = 82.8%). Inclusion of the last batch of moderators additionally resolved any differences in performance across model architectures. Time dependent analysis revealed that model generalizability for prognostic tasks has increased in more recent studies (see Supplementary material online, Figure S3).

Table 3.

Subgroup of studies for prediction task aggregated by model target and performance

First author, year of publication	Details	Sample size (P/S/V)*	Proportion women (%)	Population mean age	Network architecture	Performance metric	Value as reported
Bicuspid aortic valve
Holste, 2025		264 (24 405) P	47	69	ConvNeXt	AUC	0.83 [0.65–0.96]
Vukadinovic, 2026		NR (108 913) P	NR	NR	ViT	AUC	0.82 [0.73–0.90]
Calcification
Vukadinovic, 2026	Mitral annular	NR (108 913) P	NR	NR	ViT	AUC	0.96 [0.94–0.98]
Yuan, 2023	Coronary artery	867 (1635) P	37	72	R2 + 1D	AUC	0.85 [0.76–0.93]
Cardiac amyloidosis
Hwang, 2022	Light-chain	81 (930) P	46	51	2D CNN	AUC	1.00
Slivnick, 2025		2035 (10 278) P	53	67	R2 + 1D	AUC	0.93 [0.92–0.95]
Chao, 2024		197 (381) P	27	69	ResNet	Overall accuracy	92.7% ± 7.5%
Zhang, 2018		179 (983) V	55	57	VGG	AUC	0.87
Duffy, 2022		950 (6461) V	54	65	ResNet	AUC	0.79
Cardiac function
Holste, 2025	A) LV dysfunction B) RV dysfunction	2270 (24 405) P 3156 (24 405) P	47	69	ConvNeXt	AUC	0.99 [0.98–0.99] 0.94 [0.92–0.96]
Vukadinovic, 2026	RV dysfunction	NR (108 913) P	NR	NR	ViT	AUC	0.94 [0.88–1.00]
Liastuti, 2022	Heart failure	90 (138) P	31	50	VGG	Overall accuracy	93.0%
Holste, 2025	LV dysfunction	3298 (24 405) P	47	69	ConvNeXt	AUC	0.92 [0.90–0.93]
He, 2024	Reduced EF	152 (927) V	NR	NR	R2 + 1D	AUC	0.92 [0.89–0.94]
Cassianni, 2024	Preserved EF	646 (1284) P	NR	NR	3D CNN	Sensitivity	89.8 [87.5–92.5]
Akan, 2025	Reduced EF	NR (10 030) P	NR	NR	ViViT	Overall accuracy	87.0%
Vukadinovic, 2026	Diastolic dysfunction	NR (108 913) P	NR	NR	ViT	AUC	0.86 [0.81–0.90]
Cheng, 2022	LV dysfunction	928 (3554) V	NR	NR	3D CNN	Overall accuracy	86.0%
Akerman, 2025	Preserved EF	240 (496) P	55	75	3D CNN	AUC	0.80 [0.76–0.80]
Ramadhan, 2024	Heart failure	79 (124) V	NR	NR	ViT	Overall accuracy	77.3%
Petmezas, 2025	Reduced EF	145 (219) P	NR	70	TimeSformer	Overall accuracy	73.0% ± 0.81%
Decoodt, 2024	Heart failure	2109 (2109) P	NR	NR	3D CNN	Overall accuracy	70.0%
Shad, 2021	RV failure	182 (723) P	22	57	3D CNN	Specificity	52.7%
Cardiomyopathies
Liu, 2023	Dilated	310 (1276) P	29	52	InceptionV3	AUC	1.00
Oikonomou, 2025	Transthyretin amyloid	8842 (290 657) V	49	69	ResNet	AUC	0.99 [0.99–1.00]
Liu, 2023	Hypertrophic	121 (1276) P	29	52	InceptionV3	AUC	0.99
Hwang, 2022	Hypertrophic	191 (930) P	46	51	2D CNN	AUC	0.98
Almandani, 2024	Hypertrophic	116 (233) V	NR	NR	SlowFast + I3D	AUC	0.98
Hwang, 2022	Hypertensive	112 (930) P	46	51	2D CNN	AUC	0.96
Oikonomou, 2025	Transthyretin amyloid	112 (984) P	44	74	ResNet	AUC	0.93 [0.90–0.96]
Zhang, 2018	Hypertrophic	495 (2739) V	55	57	VGG	AUC	0.93
Chen, 2023	Hypertensive	185 (297) P	50	43	3D CNN	Overall accuracy	91.9%
Duffy, 2022	Hypertrophic	2344 (6461) P	54	65	ResNet	AUC	0.89
Oikonomou, 2025	Hypertrophic	46 398 (290 657) V	49	69	ResNet	AUC	0.89 [0.84–0.94]
Velmulapalli, 2025	Hypertrophic	1601 (8704) P	53	59	R2 + 1D	AUC	0.87 [0.81–0.91]
Akita, 2024	Hypertrophic	160 (323) P	41	55	DCNN	AUC	0.86 [0.76–0.96]
Holste, 2024	Hypertrophic	2280 (9122) P	49	69	ResNet	AUC	0.81 [0.79–0.83]
Laumer, 2022	Takotsubo	114 (228) P	92	68	Auto-encoder	Overall accuracy	75.5% ± 2.7%
Congenital heart defects
Kempny, 2019	A. TGA B. Congenitally corrected TGA	A. 92 (132) P B. 40 (132) P	41	42	2D CNN	Overall accuracy	A. 100% B. 100%
Liu, 2023	Atrial septal defect	113 (1276) P	29	52	InceptionV3	AUC	0.97
Dilation
Vukadinovic, 2026	Aortic root	NR (108 913) P	NR	NR	ViT	AUC	0.97 [0.94–0.99]
Diller, 2022	RV	308 (825) P	72	51	2D CNN	Overall accuracy	92.3%
Vukadinovic, 2026	A) IVC B) RV C) RA D) LA	NR (108 913) P	NR	NR	ViT	AUC	0.86 [0.81–0.91] 0.85 [0.80–0.90] 0.77 [0.74–0.81] 0.73 [0.70–0.75]
Medical devices
Vukadinovic, 2026	A. MitraClip B. Transcatheter aortic valve replacement C. Pacemaker	NR (108 913) P	NR	NR	ViT	AUC	A. 0.98 [0.94–1.00] B. 0.97 [0.90–1.00] C. 0.84 [0.81–0.87]
Christensen, 2024	A. Transvenous aortic valve replacement B. MitraClip C. Pacemaker	NR (224 685) P	43	66	ConvNeXt	AUC	A. 0.92 [0.91–0.92] B. 0.97 [0.97–0.97] C. 0.84 [0.84–0.84]
Ghorbani, 2020	Pacemaker	431 (3223) P	48	61	ResNet	AUC	0.89
Vafaeezadeh, 2021	Prosthetic mitral valve	447 (2044) P	NR	NR	EfficientNet	AUC	1.00
Mortality
Valsaraj, 2023	A. 1-year B. 3-year C. 5-year	A. 539 (3626) P B. 839 (2937) P C. 958 (2252) P	35	66	ResNet	Overall accuracy	A. 82.0% B. 81.0% C. 85.0%
Mou, 2026	Pulmonary hypertension patients	278 (586) P	47	61	ResNet	AUC	0.82 [0.74–0.90]
Cerna, 2021	1-year	6132 (42 095) P	NR	NR	3D CNN	AUC	0.76 [0.74; 0.77]
Myocardial infarction
Hamila, 2022		NR (162) V	NR	NR	3D CNN	Overall accuracy	90.9%
Muraki, 2022	Acute	99 (202) V	31	68	VGG	Overall accuracy	85.1%
Laumer, 2022	Acute	114 (228) P	92	68	Auto-encoder	Overall accuracy	74.1% ± 3.5%
Liu, 2023	Prior	406 (1276) P	29	52	InceptionV3	AUC	0.99
Pericardial effusion
Holste, 2025		1027 (24 405) P	47	69	ConvNeXt	AUC	0.94 [0.89–0.98]
Potter, 2023		37 (76) V	42	54	YoloV3	Overall accuracy	91.0% [73.0–99.0]
Vukadinovic, 2026		NR (108 913) P	NR	NR	ViT	AUC	0.89 [0.77–0.90]
Pulmonary arterial hypertension
Diller, 2022		450 (825) P	72	51	2D CNN	Overall accuracy	100%
Zhang, 2018		584 (3071) V	55	57	VGG	AUC	0.85
Regional wall motion abnormalities
Holste, 2025		2195 (24 405) P	47	69	ConvNeXt	AUC	0.98 [0.97–0.99]
Gul, 2026		91 (153) P	NR	NR	ResNet	Overall accuracy	94.5%
Slivnick, 2024		10 722 (25 529) S	52	71	3D CNN	Overall accuracy	92.0% [87.0–96.0]
Sahashi, 2025		678 (1453) P	42	56	R2 + 1D	AUC	0.87 [0.82–0.92]
Sanjeevi, 2023		47 (97) V	NR	NR	3D CNN	Precision	0.82
Huang, 2020		576 (5772) V	49	64	DenseNet	Overall accuracy	81.6% [80.4–82.8]
Sun, 2023		55 (137) P	NR	NR	ResNet	Overall accuracy	76.2% ± 0.06%
Stenosis
Holste, 2025	A) Mitral B) Aortic	901 (24 405) P 2020 (24 405) P	47	69	ConvNeXt	AUC	A) 1.00 [1.00–1.00] B) 1.00 [0.99–1.00]
Park, 2025	Aortic	6463 (8427) P	50	68	R2 + 1D	Overall accuracy	97.0%
Yang, 2022	Aortic	499 (3491) S	43	69	S3D	Overall accuracy	96.0% [95.0–97.0]
Vukadinovic, 2026	Aortic	NR (108 913) P	NR	NR	ViT	AUC	0.96 [0.92–0.99]
Holste, 2023	Aortic	1213 (5749) P	48	70	ResNet	AUC	0.95 [0.94; 0.96]
Holste, 2024	Aortic	1609 (9122) P	49	69	ResNet	AUC	0.95 [0.88–0.99]
Yang, 2022	Mitral	219 (3491) S	43	69	HRNet	Overall accuracy	94.0% [93.0–96.0]
Vukadinovic, 2026	Mitral	NR (108 913) P	NR	NR	ViT	AUC	0.92 [0.86–0.98]
Dai, 2023	Aortic	(16 066) P	42	66	ResNet	AUC	0.88 ± 0.01
Ahmadi, 2024	Aortic	(2247) P	NR	NR	ResNet	Overall accuracy	83.8%
Gu, 2025	Aortic	NR (2572) S	NR	59	R2 + 1D	Overall accuracy	79.7%
Vafaeezadeh, 2023	Mitral	1349 (1773) P	NR	NR	ViViT	Overall accuracy	64.1%
Duffy, 2022	Aortic	1061 (6461) P	54	65	ResNet	PR-AUC	0.08
Oikonomou, 2024	Progression—aortic	4452 (12 599) P	47	70	ResNet	AUC	0.07 [0.05–0.09]
Valve regurgitation
Holste, 2025	A) Mitral B) Tricuspid	4612 (24 405) P 4570 (24 405) P	47	69	ConvNeXt	AUC	A) 0.92 [0.91–0.94] B) 0.91 [0.90–0.93]
Vukadinovic, 2026	A) Mitral B) Aortic C) Tricuspid	NR (108 913) P	NR	NR	ViT	AUC	A) 0.91 [0.89–0.93] B) 0.89 [0.80–0.97] C) 0.88 [0.86–0.91]
Long, 2025	Aortic	9196 (52 672) S	52	61	R2 + 1D	Overall accuracy	89.3%
Yang, 2022	Aortic	434 (3491) S	43	69	U-Net	Overall accuracy	88.0% [0.86–0.90]
Holste, 2025	Atrial	1917 (24 405) P	47	69	ConvNeXt	AUC	0.85 [0.81–0.89]
Cheng, 2022	Aortic	738 (3554) V	NR	NR	3D CNN	Overall accuracy	83.0%
Yang, 2022	Mitral	928 (3491) S	43	69	U-Net	Overall accuracy	81.0% [0.79–0.83]
Long, 2025	A) Mitral B) Tricuspid	19 110 (52 672) S 21 084 (52 672) S	52	61	R2 + 1D	Overall accuracy	A) 79.7% B) 75.5%
Li, 2026	Transcatheter edge-to-edge repair	194 (633) P	43	64	ViViT	Overall accuracy	73.3% [66.3–80.4]
Vrudhula, 2024	Mitral	47 576 (75 523) V	43	NR	R2 + 1D	Overall accuracy	63.1%
Vrudhula, 2025	Tricuspid	57 867 (478 271) V	44	NR	R2 + 1D	Overall accuracy	62.6%
Valve prolapse
Al-Alusi, 2026	Mitral	783 (16 902) P	38	61	MoViNet	AUC	0.97 [0.95–0.99]
Wu, 2024	Mitral	183 (473) P	NR	NR	ResNet	Overall accuracy	94.0%
Single time
Vukadinovic, 2026	A) Elevated left ventricular pressure B) Elevated RV pressure	NR (108 913) P	NR	NR	ViT	AUC	A) 0.94 [0.88–0.99] B) 0.80 [0.78–0.82]
Park, 2025	LV outflow tract obstruction	290 (722) P	34	65	R2 + 1D	AUC	0.93 [0.92–0.95]
Holste, 2025	A) Increased LV size B) Increased LV wall thickness C) Increased LA size D) Increased RA size	1553 (24 405) P 3575 (24 405) P 6190 (24 405) P 3726 (24 405) P	47	69	ConvNeXt	AUC	A) 0.91 [0.88–0.93] B) 0.89 [0.88–0.94] C) 0.90 [0.88–0.92] D) 0.75 [0.72–0.78]
Huang, 2025	Atrial high rate episode	NR (117) P	60	76	ResU-Net	AUC	0.85
Ghorbani, 2020	A) LA dilation B) Left ventricular hypertrophy	566 (3223) P 1091 (3223) P	48	61	ResNet	AUC	A) 0.86 B) 0.75
Katsushika, 2021	Cardiac sarcoidosis	50 (199) P	54	65	R2 + 1D	Overall accuracy	70.7% [54.5–83.9]
Sahashi, 2025	Myocardial scarring	614 (1453) P	42	56	R2 + 1D	AUC	0.7 [0.61–0.78]
Chao, 2024	Constrictive pericarditis	184 (381) P	27	69	ResNet	Overall accuracy	62.1% ± 21.5%

Performance of subgroup for prediction tasks aggregated by model targets, including further details on specific diseases or subcategories of a disease as disclosed by the authors. Sample size is reported as either P, S, or V (participants, studies, or videos, respectively) and values shown are the number of individuals with the specific target and the total sample size for that study in parenthesis.

ACC, overall accuracy; CCTGA, congenitally corrected transposition of the great arteries; HF, heart failure; LA, left atrium; LV, left ventricle; MV, mitral valve; RA, right atrium; rEF, reduced ejection fraction; RV, right ventricle; TAVR, transcatheter aortic valve replacement; TGA, transposition of the great arteries; IVC, inferior vena cava.

Segmentation of cardiac structures

The overall performance of segmentation models were an external validation was performed (n = 21) can be seen in Figure 4, where considerable between-study heterogeneity (I² = 98.2%) was observed. Subgroup analysis by segmentation target did not reduce heterogeneity (I²_{left atrium} = 97.5%; I²_{left ventricle} = 94.9%), which was also the case with the addition of moderators for the basic architecture, the area of the segmented echocardiogram and number of participants used in model training (I² = 98.8%). A test of moderators revealed that the included moderators did not explain any variability in the outcome (P for model comparison=0.99). There were no differences in model performance based on the area of the echocardiogram being segmented or the DL architecture applied after the inclusion of input-related moderators. A meta-analysis showing the performance of all studies, including those that did not test their method on an external dataset, is available in Supplementary material online, Figure S4. Longitudinal assessment of Dice coefficients for the four major cardiac chambers indicates that external performance of DL segmentation models has decreased over the past seven years (see Supplementary material online, Figure S5). Meta-analysis for segmentation of the mitral valve and an overview of tasks performed only once is available in Supplementary material online, Figure S6.

Figure 4.

Meta-analysis of the dice coefficient as a measure of performance for segmentation models stratified by cardiac chamber. DC, dice coefficient; CI, confidence interval. Proportions were calculated using the testing dataset. Sample size for all studies was provided in participants or transthoracic echocardiography studies. Studies reporting sample size in echocardiographic studies were included only when it could be reasonably interpreted from the text that a one-to-one correspondence with participants was intended. Such studies are marked with an *.

Replication of established measurements in echocardiography

In studies that assessed model accuracy for LVEF prediction using the mean absolute error (%, n_studies=9), a multilevel meta-regression analysis with the basic model architecture as moderator showed high heterogeneity (I² = 98.1%). There were no significant differences in the performances of the different architectures (VGG, ConvNeXt, R2 + 1D, 3D CNN, U-Net, and PWC-Net) applied in these DL models. Meta-regression analysis showed no improved performance based on target variable. There were no significant differences observed between studies timing end-systole and end-diastole, nor between studies predicting volumes of different cardiac chambers (see Supplementary material online, Figure S7). For comparable studies, time dependent analysis of studies estimating LVEF demonstrates progressive gains in both internal and external validation performance (see Supplementary material online, Figure S8).

Classification of major cardiac imaging views

A meta-analysis of proportions was performed for those studies where the overall classification accuracy was available for individual cardiac chamber views and the models were externally validated (n = 9). The model showed high heterogeneity when classification of all chambers was analysed (I² = 75.8%). Separating the analysis into individual chambers did not resolve the high heterogeneity (Figure 5). There were no differences due to target cardiac chamber or architecture employed. A meta-analysis including all studies, internally or externally validated, is available in Supplementary material online, Figure S9. Over time, models designed for chamber view classification show enhanced generalization performance, despite a gradual decline in performance within derivation cohorts (see Supplementary material online, Figure S10).

Figure 5.

Meta-analysis of the overall accuracy for classification of cardiac chamber views. Acc, accuracy; CI, confidence interval. Proportion of correct classifications were calculated with the test dataset. Sample size for all studies was provided in participants.

Assessment of echocardiogram video quality

There was a low number of comparable studies available in quality assessment (n = 4). However, meta-analysis was still performed and it showed no differences in the performance of different architectures (MobileNetV2, ResNet, and DenseNet). Overall, accuracy in quality assessment was lower than accuracy in studies focusing on replication of established measures or disease prediction. Additionally, there was high heterogeneity that could not be further resolved (Figure 6).

Figure 6.

Meta-analysis of the overall agreement of the fully automated video-based deep learning models with human measures of quality. Acc, accuracy; CI, confidence interval. Proportion of classification of quality in agreement with expert assessment was calculated using the test dataset. Sample size for all studies was provided in participants. Studies marked with an * are those were accuracy was estimated based on the provided individual accuracies per view.

Sensitivity analyses and risk of publication bias

In studies predicting death or mortality that used accuracy as their main performance metric, there were no instances of Cook’s D above 1. Modelling of replication studies showed one instance of Cook’s D > 1, after removal heterogeneity could not be calculated due to low power. Five studies focusing on segmentation tasks had a Cook’s D above the threshold of 1. After removal, heterogeneity was fully resolved (I² = 0.00%). There were no chamber view classification studies or quality assessment studies with a Cook’s D over 1. Funnel plots assessing potential small-study effects are provided in Supplementary material online, Figure S11 and suggest that observed study effects are largely consistent with what would be expected under sampling variation alone, without evidence of substantial asymmetry. Because conventional funnel plots are difficult to interpret in diagnostic and prediction-model studies with heterogeneous metrics, tasks, and validation designs, these analyses were considered exploratory and were not interpreted as definitive evidence for or against publication bias.

Discussion

This systematic review and meta-analysis examined the current literature regarding the application of fully automated DL models to echocardiograms. To the knowledge of the authors, and within the scope of the pre-defined inclusion criteria, this is the first meta-analysis focusing specifically on fully automated, video-based analysis of echocardiography using DL. The primary focus of the analyses was the quantification of the performance of these models and the evaluation of factors influencing their accuracy. Inclusion of a large number of studies enabled a detailed investigation of various non-comparable tasks like prediction, replication, segmentation, classification, and quality assessment separately. The results demonstrated that DL models consistently achieve a good accuracy across a range of echocardiographic tasks. However, there exists substantial heterogeneity in performance that could not be resolved. This variability likely reflects systematic differences across studies, including variation in datasets, annotation standards, vendor-specific imaging characteristics, pre-processing pipelines, validation strategies, and reporting practices. In addition, publication bias, whereby studies with significant findings are more likely to be published, may have resulted in an overestimation of the effectiveness of certain models. For specific tasks, such as the segmentation of the right ventricle, for which fewer studies are available compared with more commonly assessed cardiac chambers like the left ventricle, results may be more susceptible to chance or variability within the data.²⁶

Input to a DL network, constrained to number of frames, and echocardiogram resolution, and was found to have no significant influence on model performance when examined in isolation across the studies included in this review. This observation could have several explanations. First, models trained on smaller datasets may have been better optimized to handle what are conventionally considered ‘lower quality’ inputs, thereby mitigating performance differences.²⁷ However, it is also possible that these models overfit the limited data, resulting in over-optimistic performance estimates. Second, current hardware limitations may have prevented the full use of larger or higher-quality inputs, leading to a trade-off in other computationally intensive features that enhance model performance.²⁸ Third, there may be a lack of studies exploring the extremes of input size or quality. Compounding this issue, a considerable number of studies failed to report detailed input specifications, making it difficult to fully assess their impact. This lack of transparency was likely driven by the absence of standardized reporting requirements at the time of publishing for the articles. Notably, a temporal trend was observed whereby more recent publications demonstrated substantially improved reporting quality and lower assessed risk of bias. This improvement is likely attributable to the growing adoption of structured reporting and methodological guidance frameworks, including TRIPOD-AI, PRIME 2.0, and the TRUE-AIM report card. As adherence to such standards becomes more widespread, it is anticipated that methodological heterogeneity across studies will continue to decline over time.^29–31

The studies included in this review employed two primary methodological approaches for LVEF estimation: one involving LV segmentation followed by computation using LVEDV and LVESV, consistent with standard clinical practice,³² and another based on direct regression of LVEF values from echocardiographic video data without intermediate volumetric analysis. Although the scope of meta-analysis was limited by heterogeneity in reporting, a consistent observation across multiple investigations is that DL-based measurements demonstrate lower variability compared with human interpretation.^33,34 In contrast to the well-documented interobserver variability among expert readers in routine echocardiographic assessment,³⁵ several studies have shown that automated DL approaches yield more reproducible estimates, in some cases with variability equal to or lower than that observed between experienced clinicians.³⁶ This improved consistency represents a potentially meaningful advantage of DL systems.

The findings of this systematic review should be interpreted in the context of the current landscape in which several DL-based echocardiography tools have already achieved regulatory approval,³⁷ including systems such as Caption Guidance Software (Caption Health, Inc., USA) for automated image acquisition and view guidance, EchoGo® Core (Ultromics Ltd., United Kingdom) for automated quantification, and Us2.ai (Us2.ai, Singapore) for comprehensive chamber and functional assessment. In contrast to many studies included in this review, which could only be validated in internal cohorts, these commercial systems have undergone extensive multicentre validation across diverse devices and populations to meet stringent regulatory requirements. Evidence from both commercial implementations and the studies analysed in this review suggests that DL applications for measurement automation and functional assessment can achieve reproducible performance with limited marginal gains from further incremental optimization. Accordingly, these domains appear to represent realistic opportunities for clinical adoption, provided that deployment is accompanied by appropriate workflow integration and post-implementation performance monitoring.

As these DL systems progress towards broader clinical integration, evaluation frameworks should then shift to prioritize endpoints that capture real-world utility. Rather than focusing solely on improvements in measurement precision, models should be assessed on their ability to influence patient management, streamline clinical decision making, and enhance workflow efficiency, such as reducing time to diagnosis or treatment initiation. Such outcome-oriented evaluation is essential to determine whether DL tools provide tangible benefits within healthcare systems and to guide development towards applications with meaningful clinical impact.

Despite considerable promise, integrating DL-based tools into routine echocardiographic workflows still faces considerable challenges. One fundamental barrier is the opacity of many DL models; their ‘black box’ nature can limit clinician trust and acceptance when explanations for specific outputs are not readily interpretable, hindering routine use in high-stakes decision making.³⁸ There are also ethical, regulatory, and legal considerations to take into account, ranging from data governance and bias mitigation to liability for AI-assisted decisions, which compound the difficulty of deploying DL tools at scale, emphasizing the need for structured governance and real-world validation prior to broad clinical adoption.³⁹

Although DL applications used to improve measurement acquisition appear sufficiently mature for focused clinical implementation, there are still domains where current approaches remain immature. For example, tasks such as diagnosing stenosis was found to be relatively difficult compared with other conditions. This is consistent with established methods in clinical practice, where accurate diagnosis of the condition typically relies on the combination of 2D and Doppler imaging.^40–42 The DL models in these studies, however, did not have access to the latter. These findings underscore that while DL has demonstrated clear promise in measurement automation and function assessment, substantial methodological and clinical challenges remain in domains requiring richer multimodal data or deeper pathophysiological inference.

Contemporary advances in DL are increasingly driven by transformer-based architectures and large-scale foundation models; however, their adoption within echocardiography remains limited. Among the 160 full-text articles included in this review, only 16 employed state-of-the-art architectures to develop DL-based tools with potential for physician assistance, revealing a marked methodological gap between echocardiography research and the broader DL literature. This lag suggests that many current echocardiographic applications have not yet capitalized on recent advances in model design, scalability, and representation learning. However, recent work has increasingly explored multi-task DL frameworks as one strategy to more fully leverage the flexibility of DL.^43–45 In contrast to the predominantly single-task approaches represented in our analysis, these models need to be trained on very large quantities of data and are designed to learn shared representations, potentially improving efficiency and consistency within a unified pipeline. Such approaches may offer one avenue to mitigate some of the heterogeneity observed across studies. While still emerging and not yet sufficiently represented to influence the quantitative results of this work, these developments signal a methodological broadening of the field that may shape future performance and validation patterns.

Most of the DL research in echocardiography focuses on incremental performance improvements for tasks that are advanced and technically mature and, in some cases, approved by regulatory authorities. While such work remains important for validation, robustness, and generalizability, the predominance of studies revisiting similar tasks―often using overlapping datasets and comparable model architectures―suggests that further methodological gains may yield diminishing returns in terms of incremental clinical benefit. In contrast, only few studies have focused on underexplored disease entities, longitudinal risk stratification, or decision-critical clinical questions where DL-based approaches might address unmet needs in echocardiographic practice. Furthermore, incorporation of multiple views remains limited. Shifting a greater proportion of research effort towards these higher-impact domains may better align methodological advances with clinically meaningful outcomes and help bridge the gap between technical performance and real-world utility.⁴⁶

While DL holds clear promise to reduce physician workload and to augment clinical interpretation, most existing applications remain confined to tasks that replicate established human assessments. However, DL methods are uniquely capable of identifying complex, high-dimensional patterns and temporal dynamics that may be imperceptible to human observers, enabling the detection of subtle phenotypic signatures of disease. Only a comparatively small number of studies have explored prognostic modelling or diagnostic tasks not routinely addressed using echocardiography; as a result, the rich information embedded in echocardiographic videos remains largely underutilized.

Evidence from other imaging domains illustrates the potential of this paradigm. For example, the MIRAI DL model developed for mammography demonstrated the ability to predict an individual’s risk of developing breast cancer up to five years in advance.⁴⁷ Such findings suggest that medical imaging modalities contain relevant information beyond current human interpretation. In echocardiography, similar DL approaches may therefore enable detection of cardiovascular conditions and risk patterns not typically captured by standard clinical assessment. A small subset of studies included in this systematic review have already begun to explore such applications, including prognostic modelling and non-traditional tasks derived directly from echocardiographic data. However, these efforts remain the exception rather than the norm. This imbalance highlights both the early emergence of more exploratory, insight-driven applications and the largely untapped opportunity for future research to fully exploit the unique strengths of DL.

Strengths and limitations

Several limitations should be acknowledged. The exclusion of non-peer reviewed work meant that studies available only on preprint servers such as arXiv, where state-of-the-art developments in computer science are often reported, were not included in this systematic review. The methodological heterogeneity of included studies may have introduced publication bias. Although the application of Cook’s D helped to identify and mitigate the influence of outliers, inconsistencies in reporting remain a significant concern. Performance metrics for identical prediction targets were often reported using non-standardized formats, and in many cases, the absence of measures of variability, such as standard deviations or confidence intervals, precluded inclusion in the meta-analysis. Additionally, overlap in training datasets, derivative model development, and repeated use of the same validation cohorts across studies may have reduced the independence of some performance estimates and should be considered when interpreting the pooled results. Certain moderator analyses could not be conducted due to incomplete or insufficient reporting of key study characteristics. Importantly, this review is restricted to published studies and does not include proprietary or commercial DL software. The performance of these systems may more closely reflect the true achievable performance of DL in echocardiography, given the rigorous validation required to meet regulatory standards. Their exclusion represents a potential source of bias and should be considered when interpreting the generalizability and clinical applicability of the pooled estimates. Additionally, limited reporting on algorithm retraining constrains full assessment of model performance and generalizability.

Despite these challenges, which are consistent with previously reported limitations in applied DL research,⁴⁸ a considerable number of studies met the inclusion criteria, enabling valuable insights into the functioning of DL models as automated tools in echocardiography. The formation of subcategories based on prediction tasks and model architectures facilitated comparative analyses and yielded clinically and technically relevant insights into the performance of DL models as fully automated tools in echocardiographic assessment.

Conclusion

This systematic review and meta-analysis demonstrate a growing and increasingly robust body of evidence supporting the application of DL in echocardiography across a broad spectrum of clinical tasks. The available evidence suggests that some of these applications may be approaching readiness for clinical integration. If successfully validated and implemented, they have the potential to improve consistency, reduce interobserver variability, and alleviate physician workload, particularly in resource-limited settings, while complementing rather than replacing clinical judgment. Greater research effort should be directed towards unexplored areas, including longitudinal risk stratification, multimodal approaches, and diagnostic tasks not traditionally addressed by echocardiography, where DL’s capacity to detect subtle patterns may unlock genuinely novel clinical insight. Realizing this potential will require prioritization of regulatory readiness, external validation, and transparent model interpretability to foster clinician trust and facilitate responsible clinical integration. If these conditions are met, DL stands to move beyond replicating human expertise and towards augmenting it in ways that meaningfully advance cardiovascular care.

Supplementary material

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

Author contributions

P.S.W., P.L., S.K., and S.E. acquired funding for the study. A.R.M., V.t.C., S.Z., P.L., S.E., S.K., S.-O.T., and P.S.W. conceived and designed the study. P.S.W. supervised the work. A.R.M., S.Z., M.K., V.t.C., and K.K.B. conceived the analysis. A.R.M., S.Z., M.K., and K.K.B. conducted the analysis. A.R.M., S.-O.T., and P.S.W. drafted the manuscript. All authors critically reviewed the manuscript and approved the final draft.

Funding

The present work and A.R.-M., K.K.-B., and V.t.C. were funded by the German Federal Ministry of Research, Technology, and Space (BMFTR 03ZU1202LB, 03ZU1202JC, 03ZU1202AA, respectively) within the future cluster curATime. S.Z. was funded by the Carl-Zeiss-Stiftung within Multi-dimensionAI (CZS-Project number: P2022-08-010). P.S. Wild is principal investigator of the German Centre for Cardiovascular Research (DZHK), principal investigator of the BMFTR DIASyM research core (BMFTR 161L0217A, 031L0217A, 16LW0241K), principal investigator of the future cluster curATime (BMFTR 03ZU1202AA, 03ZU1202CD, 03ZU1202DB, 03ZU1202JC, 03ZU1202KB, 03ZU1202MB, and 03ZU1202OA), principal investigator of EPIC-AI (BMFTR 01EQ2401A) and principal investigator of Multi-dimensionAI (CZS-Project number: P2022-08-010), funded by the Carl-Zeiss-Stiftung.

Data availability

Data are available upon reasonable request to the corresponding author.