Heart failure (HF) remains a significant global health problem leading to high hospitalization and mortality rate despite advances in therapy [1]. The burden of the disease in Asia is particularly more pronounced, considering that it affects a younger population than in Europe and America [2, 3]. Early detection and treatment of possible cases are mandatory to prevent disease progression and reduce health care costs.

Echocardiography is a widely recommended imaging modality for assessing cardiac function in HF patients [4, 5]. Although echocardiography is non-invasive, harmless, and relatively inexpensive, some severe issues have arisen regarding its implementation. Echocardiography test is largely dependent on the user’s skill, creating challenges for interpretation [6]. Furthermore, the terminology of left HF comprises a wide range of phenotypes, from those with systolic dysfunction or reduced ejection fraction (HFrEF) [EF 40%], diastolic dysfunction or preserved ejection fraction (HFpEF) [EF $\ge$50%], and the ‘grey area’ cases with mid-range ejection fraction (HFmrEF) [EF 40–49%] [5]. Diagnosing HFpEF from echocardiography alone is not a simple task as the European Society of Cardiology guidelines recommends combining with other diagnostic tests, including natriuretic peptides level and electrocardiogram (ECG) [5, 7].

The most potential solution to the limitation of echocardiography interpretation lies in the application of automated methods, which have vastly evolved through computer technology. Artificial intelligence leverages computers and machines to mimic the human mind in problem-solving capacities. It enables training of large databases of various echocardiographic videos and images which have been previously confirmed by experts to achieve knowledge which is then used to identify endocardial pathologies in other cases [8].

A total of 2105 citations were retrieved by the method aforementioned. After reading titles and abstracts and assessing these articles for eligibility, 2066 citations were excluded. Full-text articles were assessed, and 25 articles were excluded as these literatures did not match our PICO or were duplicate articles. As a result, 14 articles remained and included in this systematic review [6, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27]. The detailed elaboration of PRISMA flow is described in Fig. 1.

Fig. 1.

PRISMA flow diagram of study selection process.

Only a few studies pose unclear risk of bias in terms of patient selection, including Dong et al., Ghorbani et al., Behnami et al., and Liu et al. These risk of bias was caused by unclear inclusion and exclusion criteria and an unspecified randomization process. Studies with unclear risk of bias in their index test, including Dong et al., Ghorbani et al., Bhenami et al., Chiou et al., Kusunose et al., Liu et al., and Chen et al., demonstrated ambiguity of blinding process. Other studies exhibit low risk of bias and low applicability concern in all of the QUADAS-2 variables. Complete QUADAS-2 assessment of these literatures are explained in Table 1 [6, 13, 14, 15, 16, 17, 19, 20, 22, 23, 24, 25, 26, 27].

The characteristics and results of these 14 studies are summarised in Table 2 [6, 13, 14, 15, 16, 17, 19, 20, 22, 23, 24, 25, 26, 27]. Nearly all datasets are in 3D or 2D images, and the most common echocardiography views used in training are apical four-chamber (A4C) and apical two-chamber (A2C). Supervised ML is frequently used as the algorithm’s core, notably convolutional neural network, to classify and differentiate individuals with and without the disease.

Dong et al. [13] proposed a method incorporating unsupervised multi-scale convolutional deep networks and random forests to predict LV volume and calculate LVEF. The multi-scale convolution deep network extracted features of unlabelled end-diastolic and end-systolic 3DE volumes (EDV and ESV). Afterward, the left ventricular volume was formulated as a regression problem; thus, random forests were used to estimate the efficient volume.

Raghavendra et al. [14] developed a framework of ML techniques to classify CHF due to dilated cardiomyopathy and normal controls. The 2-D images were decomposed to generate specific structural patterns of each group using variational image decomposition (VMD). After the texture feature is extracted and enhanced using particle swarm optimization (PSO), the support vector machine (SVM) separates the class members into two groups.

Sanchez-Martinez et al. [15] combined unsupervised ML algorithms to investigate left ventricular long-axis myocardial velocity patterns that ordered subjects according to their similarities, allowing further analysis of the main trends in velocity patterns. The clustering system identified a continuum from normal to HF, including a transition zone of uncertain diagnosis. This method was subsequently independently validated in two additional cohorts, breathless and hypertensive patients. These resulted in limited accuracy and misclassification into the HFpEF group.

Almost all recent studies adopted CNN (convolutional neural networks) as the principal classifier, but each has modifications to elevate diagnostic power. Silva et al. [16] demonstrated custom 3D-CNN ability to integrate temporal knowledge from transthoracic echocardiography (TTE) cine loops to calculate LVEF and classified it into four classes. Ouyang et al. [19] and Ghorbani et al. [20] presented a novel CNN-based ML called Echo-Net. The algorithm performed several tasks, from left ventricle segmentation during systole and diastole, beat-to-beat prediction of the ejection fraction, and presence of heart failure conclusion. The ML was also able to identify the local cardiac structures and anomalies, measure volumetric parameters and metrics of cardiac function, and predict systemic human phenotypes that modify cardiovascular risk. Behnami et al. [22] built combined supervised ML for binary EF classification without segmentation. What is even more interesting is that two image views, A4C and A2C, were concatenated for temporal embedding. Kusunose et al. [24] compared two types of input method averaged images and ten selected images from 5 standard views and tested them using 3D-CNN to recognize the view type and estimate LVEF. The group with selected images improved the overall accuracy of echo-view classification and LVEF estimation.

Several studies carried out diagnostic classification after image segmentation using U-net. Chiou et al. [23] used U-net for left atrium and left ventricle segmentation to measure their length, width, area, and volume. The interbeat dynamic changes were then recorded as linear waveform signals, trained and classified by a 1D CNN. Liu et al. [25] presented a DPS-Net model, a constructed CNN based on modified U-net, and tested the ML on a local dataset of A2C, A3C, and A4C images. After that, the algorithm was retrained using a sizeable multicenter dataset to generate better accuracy in view classification, end-systolic and end-diastolic frame detection, and by all means, LVEF measurement.

ML analyses using datasets of the Doppler modality have recently become more promising. Tabassian et al. [17] investigated spatiotemporal characteristics of velocity, strain, and strain rate traces during rest and exercise from tissue Doppler using a supervised and unsupervised algorithm. Each parameter of the rest and exercise tests was concatenated, and the pattern was analyzed using principal component analysis (PCA). Further, automatic classification using distance-weighted ĸ-nearest-neighbor (DWKNN) was applied to differentiate HFpEF cases and multi-phenotype controls. Pandey et al. [26] developed a combination of unsupervised and supervised learning and trained the algorithm using the dataset of routinely measured Doppler indexes. The model was also implemented in the hemodynamic external validation cohort to identify two phenogroups (high-risk vs. low-risk) patients and demonstrated a strong diagnostic value. Chen et al. [27] tested the performance of Deep CNN, which was predicted to have better application recognition performance due to more layers in the ML architecture and simplified connection. Aside from 2D images data, Doppler indexes, including atrial systolic velocity (A), early mitral valve diastolic maximum velocity (E), were also processed to enhance the diagnostic accuracy. The latest study is from Tromp et al. [6], which presented a new ML approach of 2D videos and Doppler parameters that allows fully automated classification and annotation of echocardiographic videos. These 2D videos were classified into views by two different classifiers, a supervised CNN or an unsupervised deep clustering CNN. Meanwhile, the Doppler modalities view classifier consisted of integrated CNN models trained with the echo or velocity trace images and the categorical ground truth labels.

Our systematic review represents the model’s current state for diagnosing heart failure more rapidly through echocardiography images. The literature included in this study has shown that AI has comparable performance in characterizing heart failure through echocardiography images, compared with the conventional method by medical practitioners, with an accuracy rate ranging from 57% to 99.3%. Supervised ML, particularly CNN, was the most utilized algorithm, and few of them optimized external datasets.

Current evidence has shown that upon using sufficient training datasets, various AI approaches can bring astounding performance in many tasks, such as object-identifying tasks, the main application in medical diagnosis across the reviewed literature [28]. The assistance of AI technology can make rapid detection of clinical symptoms based on the image features, like tone and rim. Computer-assisted technology is also capable of producing consistent outputs, that will lead to increasing the efficiency of healthcare service, saving vast amounts of time in clinical practice, complementing cognitive fatigue, and markedly reducing the workload of clinical practitioners. Nevertheless, the application of AI cannot be isolated from its clinical significance.

Each algorithm has its advantages and disadvantages. The conventional ones such as random forest and support vector machines might have better interpretability and be cheaper in computation cost than the deep learning-based algorithm. With that said, having expertise to decide how the features are extracted from the data is necessary. They heavily rely on such well-defined features; hence their performance is dependent on successful feature extraction. Manual feature extraction is a tedious task; therefore, many believe it is time-consuming, labor-intensive, and inflexible [29]. On the other hand, the deep learning-based algorithm can extract the feature independently. Therefore, it does not require expertise to perform the feature extraction task manually. The feature extraction and classifier are often end-to-end connected and learn together through optimization algorithms such as gradient descent. It results in a fully automated feature extraction and model training process. Therefore, it is considered the antidote to the conventional AI algorithm drawback.

In computer vision, a convolutional neural network (CNN) is a very well-known deep learning-based algorithm designed to work with grid-structured inputs, which have solid spatial dependencies in local regions of the grid such as image and video [30]. It consists of a convolution layer and pooling layer used to extract features, like edges, corners, shapes, from the input image and feed-forward it to the next layer. Each convolution layer has its parameters that can be learned during the gradient descent process; therefore, this model does not require human expertise for feature extraction since it can do the task on its own. Moreover, the shared parameters of the filter across the entire convolution make this model be equivariance to translation. In other words, if we shift an object in an image, it does not alter the representation of the data in the deeper layer of CNN. Therefore, this model often gives a promising performance.

The drawback of a deep learning-based algorithm lies in its interpretability since the model seems to learn thoroughly on its own that it is hard to explain what the model is trying to see from the data. Moreover, it is popularly known as computationally expensive that it often requires a graphical processing unit (GPU) to run the deep learning-based algorithm since it often has millions of parameters to compute. Deep learning is tough to train; even if the model managed to have 100% accuracy on the training data, it does not guarantee that the model has the same performance for unseen data. This phenomenon is called overfitting, where the model is suitable only for the training data. The simplest solution to this problem is to collect more data. With increased training data, the training accuracy will be decreased due to more diversity in the data. However, that will make the model more general, resulting in good predictions for the unseen data. The number of samples appearing for each class must be considered to avoid imbalanced class problems. An imbalanced class problem occurs when samples from a specific class occur more frequently than others. For example, if the collected data is 1000 with 700 of which are normal while 300 of which are heart failure, the model trained with this kind of imbalanced class will be biased toward the normal class, resulting in prediction attempt more frequent to the normal class than the heart failure class. This kind of model will likely have a low recall score. That is why the number of samples from each class must be considered when collecting the data.

While the advancement of AI technology might be promising, a medical evaluation by the experts still plays a vital role. The final diagnosis of the disease shall have a real-world impact to improve patients’ health; thus, AI cannot be separated from human engagement; they must work together in harmony [31]. However, AI also poses some limitations: (1) It needs high-quality datasets for training and validation; (2) There could be ethical and safety issues, e.g., using AI after obtaining the patient’s consent and determining who is liable for a misdiagnosis or incorrect treatment; and (3) it cannot determine causal relationships; thus it still need evaluation and interpretation by medical practitioners [32].

Despite having particular challenges, the future of AI in cardiology is promising in the era of precision medicine, especially in diagnosing heart failure. Heart failure has complex pathophysiology with various clinical features; thus, its diagnosis can be challenging even for cardiologists [33]. Patients with HF can have a poor prognosis and high readmission rates. The use of AI can be beneficial to rapidly detect the disease in early stages, thus improving the patient’s prognosis and saving lives. Consequently, misdiagnosis of HF can hinder the chance of improving a patient’s outcome. AI models have the potential to make better medical decisions, reduce clinical errors, and improve quality of life [34].

Even though we have covered various major databases of scientific articles, this systematic review still has the potential weakness of only including studies from the published literature and eliminating other studies, such as conference abstracts. These conference abstracts were primarily recent studies published in 2020–2022. Unfortunately, the information in these kinds of literature did not have the full study details. We might have missed some articles, especially those published in languages other than English. This systematic review highlights the need for additional research regarding the use of AI in heart failure diagnosis. Our study is the first to review the current literature on heart failure diagnosis through echocardiography and AI.

Cardiovascular imaging, particularly echocardiography, is an essential tool for medical practitioners, especially to detect left heart failure patients as early as possible. Studies have shown that artificial intelligence has a high potential to serve as practical auxiliary assistance for medical practitioners to differentiate normal and left heart failure patients through echocardiography. It is unlikely that artificial intelligence will completely replace cardiologists in interpreting echocardiography images, diagnostic processes, and overall clinical care. Despite limitations, AI remains a vital concept in the future of cardiology, and additional research is needed.

LDL and AAN designed the research study. YN, AAN, and GJ performed the research. BBS, RS, and WJ provided help and advice on the literature search by providing grey literatures, improving the methodology and overall analysis of the study. YN and RYIP analyzed the data. YN, RYIP, and GJ wrote the manuscript. All authors contributed to editorial changes in the manuscript. All authors read and approved the final manuscript.

Not applicable.

We would like to express our gratitude to all those who helped us during the writing of this manuscript.

This research received no external funding.

The authors declare no conflict of interest.