Artificial Intelligence in Diagnosis of Heart Failure

Abstract

Heart failure (HF) is a complex and varied condition that affects over 50 million people worldwide. Although there have been significant strides in understanding the underlying mechanisms of HF, several challenges persist, particularly in the accurate diagnosis of HF. These challenges include issues related to its classification, the identification of specific phenotypes, and the assessment of disease severity. Artificial intelligence (AI) algorithms have the potential to transform HF care by enhancing clinical decision‐making processes, enabling the early detection of patients at risk for subclinical or worsening HF. By integrating and analyzing vast amounts of data with intricate multidimensional interactions, AI algorithms can provide critical insights that help physicians make more timely and informed decisions. In this review, we explore the challenges in current diagnosis of HF, basic AI concepts and common AI algorithms, and latest AI research in HF diagnosis.

Heart failure (HF) continues to pose a significant global health challenge, affecting over 50 million people worldwide. ¹ This complex and multifaceted syndrome is marked by the heart's inability to pump adequate blood to satisfy the body's metabolic demands. The diagnosis and management of HF are particularly challenging due to its diverse manifestations and the variability in patient responses. Despite substantial advancements in medical research and technology, the traditional methods for diagnosing HF often fall short, primarily due to the disease's heterogeneous nature. ² The challenges of accurate classification, phenotyping, and severity assessment hinder timely and effective intervention. Early and precise diagnosis is crucial for improving patient outcomes and managing the condition effectively. Over the course of illness, patients with HF receive numerous invasive and noninvasive diagnostic tests, generating large amounts of medical data. The size, complexity, and dynamic nature of big data can pose challenges for traditional statistical methods. In this evolving landscape, artificial intelligence (AI) offers promising new avenues for transforming HF diagnosis. ³ , ⁴ , ⁵ , ⁶ , ⁷ , ⁸ , ⁹ , ¹⁰ , ¹¹ , ¹² By leveraging advanced algorithms and machine learning techniques, AI has the potential to enhance diagnostic accuracy, facilitate earlier detection, and support clinical decision‐making. These technological innovations could significantly address the gaps in current diagnostic approaches and improve the overall management of HF, ultimately leading to better patient outcomes and more efficient health care delivery. In this review, the goals are 4‐fold: (1) to point out the challenges of diagnosing HF, (2) to delineate the general AI concepts and methodology, (3) to highlight important applications of AI in HF diagnosis, and (4) to note the overall limitations of AI.

THE CHALLENGES OF DIAGNOSING HF

Definition and Classification of HF

The proposed universal definition describes HF as a clinical syndrome with symptoms or signs caused by a structural or functional cardiac abnormality and corroborated by elevated natriuretic peptide levels or objective evidence of pulmonary or systemic congestion. Patients with HF may present with a wide range of symptoms, from fatigue and shortness of breath to fluid retention (pulmonary congestion, systemic venous congestion, and peripheral edema). This variability makes it challenging for clinicians to recognize HF early, especially in patients with atypical presentations or in those with comorbid conditions that can mask or mimic HF symptoms. Additionally, the symptoms of HF often overlap with those of other cardiovascular and noncardiovascular conditions, such as chronic obstructive pulmonary disease, renal disease, and obesity. This overlap can lead to misdiagnosis or delayed diagnosis, particularly in patients with multiple comorbidities. Differentiating HF from other conditions requires a careful and comprehensive evaluation, often involving a combination of clinical assessment, imaging, and laboratory tests. Diagnosing HF is inherently complex due to the multifaceted nature of the condition.

Left ventricular ejection fraction (LVEF) is a key factor in classifying patients with HF because it influences prognosis and treatment responses and is commonly used as a selection criterion in clinical trials based on ejection fraction. According to the 2022 American College of Cardiology and the American Heart Association guidelines, HF can be classified into 3 types based on LVEF, which are HF with preserved EF (HFpEF), HF with mildly reduced EF, and HF with reduced EF (HFrEF). ² HFrEF is defined by an LVEF ≤40%, and HFpEF is characterized by an LVEF ≥50%. Patients whose LVEF falls between these 2 ranges are described as HF with mildly reduced EF. Classification of HF is shown in Table 1.

Table 1.

Classification of HF by LVEF

Type of HF according to LVEF	Criteria
HF with reduced EF	LVEF ≤40%
HF with mildly reduced EF	LVEF 41%–49%
	Evidence of spontaneous or provokable increased LV filling pressures (eg, elevated BNP, noninvasive and invasive hemodynamic measurement)
HF with preserved EF	LVEF ≥50%
	Evidence of spontaneous or provokable increased LV filling pressures (e.g., elevated B‐type natriuretic peptide, noninvasive and invasive hemodynamic measurement)

EF indicates ejection fraction; HF, heart failure; LV, left ventricular; and LVEF, left ventricular ejection fraction.

As mentioned previously, HF is a heterogeneous entity, representing multiple phenotypes with distinct pathophysiological characteristics. Starting in 2013, several clustering studies focused on HFrEF. ¹³ , ¹⁴ , ¹⁵ , ¹⁶ Since 2015, clustering efforts have largely concentrated on HFpEF. ¹⁷ , ¹⁸ , ¹⁹ , ²⁰ , ²¹ , ²² Shah et al. implemented a range of detailed data inputs and statistical learning methods to produce 3 unique subgroups, with remarkably different clinical characteristics, cardiac structure/function, invasive hemodynamics, and outcomes. ¹⁷ Phenomapping results in a novel classification of HFpEF. Since this study, there has been a significant increase in phenomapping efforts in HFpEF, using various data elements and statistical techniques. Peters et al. reviewed insights, limitations, and future directions of phemomapping in HFpEF. ²³ Despite some heartening results, they have yet to be incorporated into clinical practice due to generalizability, single‐center design, and lack of longitudinal data.

Diagnostic Criteria and Algorithm of HF

The guidelines on diagnosis and treatment of HF have been updated over time, reflecting advances in clinicians' understanding of the disease. ² , ²⁴ , ²⁵ The European Society of Cardiology has created a diagnostic algorithm that starts with a preliminary assessment of HF symptoms and signs, common clinical demographics, and diagnostic tests including laboratory work, ECG, and echocardiography. ²⁴ Diagnostic algorithm for classification of HF according to LVEF is shown in Figure 1.

Figure 1. The algorithm for a diagnosis of EF‐based HF.

EF indicates ejection fraction; HF, heart failure; HFmrEF, heart failure with mildly reduced ejection fraction; HFpEF, heart failure with preserved ejection fraction; HFrEF, heart failure with reduced ejection fraction; LVEF, left ventricular ejection fraction; and NT‐proBNP, N‐terminal pro‐B‐type natriuretic peptide.

For HFrEF, diagnosis is relatively straightforward. An LVEF ≤40% is indicative of HFrEF. However, to distinguish between HFpEF and healthy population, difficulties remain. Several challenges need to be addressed in the diagnosis of HFpEF. First, according to 2019 European Society of Cardiology consensus recommendation, ²⁴ the diagnostic algorithm for HFpEF (heart failure association pretest assessment, echocardiography and natriuretic peptide, functional testing, and final etiology) is quite complicated, which needs the integration of multiple parameters from different testing for assessment. Additionally, the diagnostic algorithm applies exclusively for HFpEF. Second, the gold standard for diastolic filling pressure evaluation is right heart catheterization, a highly limited and invasive procedure. Lastly, even for alternative noninvasive imaging methods such as echocardiography, the assessment of LV filling pressure is quite complicated. The algorithm for diagnosis of LV diastolic dysfunction by echocardiography requires accurate measurement of average E/e', septal and lateral e', left atrial volume index, and tricuspid regurgitation velocity.

Therefore, despite a lack of inherent flaws within current diagnostic criteria or algorithms for HF, its complexity somewhat limits its application, potentially leading to underdiagnosis or misdiagnosis, especially for HFpEF.

Diagnostic Tools for HF

Echocardiography is a cornerstone in the diagnosis of HF, providing critical information about cardiac structure and function. ²⁶ , ²⁷ It is also the only imaging technique that can evaluate diastolic function. However, the accuracy of echocardiographic assessment can be influenced by operator expertise, patient body habitus, and the quality of the images obtained. ECG plays a crucial role in detecting arrythmia and myocardial infarction, a complication that many patients with HF have. ²⁶ There are many differences between ECG signals between patients with HF and healthy patients. Yet the large amount of ECG data is not fully used clinically, which could provide useful information for early diagnosis of HF. Biomarkers such as BNP (B‐type natriuretic peptide) and NT‐proBNP (N‐terminal proBNP) are commonly used for the screening and diagnosis of HF. ²⁸ , ²⁹ , ³⁰ However, their levels can be influenced by a variety of factors unrelated to HF, including age, renal function, and obesity. This variability can complicate the interpretation of biomarker levels.

Access to these diagnostic tools can vary significantly based on geographic location and health care infrastructure. In resource‐limited settings, clinicians may rely on less comprehensive diagnostic methods, leading to potential delays in diagnosis or misdiagnosis. ³¹ , ³² This disparity underscores the need for scalable, cost‐effective diagnostic methods that can be widely implemented, such as AI.

Potential Role of AI in HF

With the ability to perform tasks at human intelligence level, AI has been gaining an important role in the management of cardiovascular diseases. Advances in AI also shed light on unresolved issues in the diagnosis of HF. For HFrEF, which is relatively straightforward and easy to evaluate, AI could replicate diagnostic processes, accelerate diagnosis, and reduce clinicians' workload by automating LVEF quantification and disease detection. HFpEF, which is more difficult to diagnose and often unrecognized clinically, could benefit even more from AI methods. By accomplishing complex analysis of large heterogenous data from patients with HFpEF, AI shows great promise in terms of saving manual labor, reducing health care expenditures, improving disease detection as well as phenomapping distinct subgroups. Unsupervised machine learning (ML), a subfield of AI, has the potential to uncover new insights into HFpEF studies. It discovers new patterns between variables by grouping data through repeated iterations, thus developing novel classification systems for patients with HFpEF. Although ML has great potential to improve HF research, the applications must be validated in prospective studies. In addition, AI can incorporate multiple parameters obtained from multimodality imaging techniques, ³³ , ³⁴ , ³⁵ which may be a noninvasive surrogate for catheterization.

History, Concepts, and Performance Measures of AI

Just as the steam engine was the driving force of the age of steam, the generator of the age of electricity, and the computer and internet of the information age, AI is becoming the key force that drives humanity into the age of intelligence. To understand where AI is headed, one must first know where it came from. The journey began in the 1950s with early theorists like Alan Turing, who questioned if machines could think and proposed the Turing Test to measure machine intelligence. ³⁶ In 1956, the Dartmouth Conference, organized by John McCarthy, officially introduced the term “artificial intelligence” and laid the groundwork for AI research. ³⁷ AI has evolved significantly since its inception. The 1990s marked a resurgence in AI research. This era saw the development of algorithms that could learn from data and improve their performance over time. Notable milestones include IBM's Deep Blue defeating world chess champion Garry Kasparov in 1997. ³⁸ The 2000s saw breakthroughs in image and speech recognition, exemplified by technologies like Google's search algorithms and Apple's Siri. At the core of AI are several foundational concepts, including machine learning and deep learning (Table 2 and Figure 2). ML is a subset of AI and involves training algorithms on data to enable systems to improve their performance over time without being explicitly programmed for specific tasks. ³⁹ Supervised ML is trained on a labeled data set and unsupervised ML involves finding patterns or groupings in data without preexisting labels. ⁴⁰ Semisupervised learning is a type of ML that combines supervised and unsupervised learning by using labeled and unlabeled data to train AI. Deep learning (DL) leverages computational models composed of multiple interconnected processing layers to automatically learn and refine data representations at increasing levels of abstraction. ⁴¹ This hierarchical structure allows the model to capture and process intricate details within data, which has led to significant advancements across various domains. Natural language processing is a dynamic field of AI that focuses on the interaction between computers and human language. ⁴² , ⁴³ It involves the development of algorithms and models that enable machines to understand, interpret, and generate human language in a way that is both meaningful and useful. Natural language processing is not inherently part of DL, though it is currently dominated by approaches like recurrent neural networks and transformers. Specific performance measures for AI algorithms are summarized in Table 3.

Table 2.

Common Artificial Intelligence Algorithms

Machine learning
Algorithm	Underlying principles	Strengths	Weaknesses
Unsupervised learning
K‐means clustering	The algorithm identifies k number of centroids and computes the sum of squared Euclidean distances between the centroids and the data points on a coordinate plane. Each data point is then allocated to the closest centroid to keep the clusters as small as possible. The final centroids are computed by taking the averages of all the data points that belong to each cluster.	+ Relatively simple and computationally fast to implement + Flexible approaches with reasonable results for most problems	– Identifying the number of clusters can be challenging – Accuracy decreases if clusters are nonglobular or vary in size and density
Mean shift	Mean shift begins with a circular sliding window centered on a randomly selected data point, with a predefined radius. At each subsequent iteration, the window is gradually shifted toward the region of higher point density. The algorithm continues until there is no furthur a shift that can increase the point density within the window.	+ No need to select the number of clusters	– The selection of window size can be challenging
Expectation maximization	The algorithm begins by randomly assigning initial parameters: mean±SD based on the user‐defined number of clusters. The algorithm then refines the parameters through an iteration of two steps: the expectation (E) step and the maximization (M) step. In the E step, the algorithm calculates the probability of each data point belonging to a cluster based on the initial parameter values. In the M step, a new set of parameters is computed to maximize the likelihood of the distribution provided in the E step. At the end of the process, each data point is assigned to its most probable cluster.	+ Can identify nonglobular clusters + Allow for the extraction of membership probability for each data point	– The number of clusters must by identified by the user – Can be much slower than other algorithms
Supervised learning
Logistic regression	Logistic regression is a statistical technique used to estimate the probability of a binary response using logistic/sigmoid function. Logistic regression can handle multiclass classification by using the one‐vs‐rest scheme, in which each class is treated as a binary classification problem.	+ Interpretable + Computationally efficient and easy to implement + Does not require scaling of input features + Provides the probability score for the observations	– Vulnerable to overfitting – Unable to handle data set with large number of categorical inputs – Collinearity between input variables negatively affects the accuracy
K‐nearest neighbors	Assumes that similar data points are in close proximity to each other. The algorithm classifies a data point based on the class of the nearest neighbors. The number of the neighbors to be used is predefined by the user. The majority class of the neighbors is taken to be the prediction of the query instance.	+ Simple algorithm to implement and interpret. + Quick computing time + No major assumptions about data	– User must define the number of nearest neighbors to use – Memory intensive – Poor performance with high‐dimensional data
Support vector machine	Checks for a hyperplane (or a line in 2‐dimensional data) that best separate the 2 different classes. The best hyperplane is defined as the one that maximizes the distance to the nearest element of each class.	+ Works well when there is a clear boundary between classes. + Memory efficient.	– Low accuracy when samples overlap – Does not work well with large data set – Does not provide probability estimates
Decision tree	A decision tree starts with the best feature as the root node and recursively splits the data into subsets until each subset has the same classification. If a subset contains multiple classification and no features remain to split, the algorithm returns the most frequent class. The decision of making strategic splits is based on whether a split will increase the homogeneity of the resultant subsets, which can be calculated using the Gini index.	+ Closely mimics the human decision‐making process + Easy interpretation + Work well with both numerical and categorical features	– Prone to overfitting – Affected by noise
Random forest	Random forests are an ensemble learning algorithms, which is a combination of multiple base decision tree. Each base decision tree is trained on a subset of the initial training set, using bootstrapped samples. The final classification of a data point is based on the simple majority voting scheme.	+ Resilient to overfitting. + Works well with nonlinear data + Robust to outliers + No scaling required.	– Takes long to train. – Low interpretability
Boosting algorithms (XGBoost, Adaboost, Catboost, LightGBM)	Boosting algorithms are also a form of ensemble learning. Instead of creating several base models independently, like random forests, boosting algorithms create the base decision tree sequentially. Each successive tree focuses on the errors of the previous trees by assigning higher weights to the misclassified instances, while keeping the weights of the correctly classified instances the same. The final model is based on the combination of all weighted base models.	+ Easy to interpret + Resilient to overfitting + No scaling required + Good performance + Work well with both numerical and categorical features.	– Slow to train since the model is built consequentially
Deep learning
Neural networks	A neural network consists of neurons, which are the basic processing units of the network. Each neuron transform the input data with a weighted linear summation, followed by a nonlinear activation function. The neurons are organized into several layers, where the output of one layer is used as the input for the next layer. Each layer can thus identify important features from the input the previous layer and further process them.	+ Allows for capture of complex and nonlinear relationships from the data + Excellent performance with large data sets	– Require large amounts of data – Expensive to compute – Black boxes nature – Heavily depend on the quality of the training data – More challenging than other algorithms to design and develop
Natural language processing
Symbolic NLP algorithms	Symbolic algorithms, also known as rule‐based or knowledge‐based algorithms, rely on predefined linguistic rules and knowledge representations.	+ Highly interpretable + Can handle complex linguistic structures + Effective for specific tasks where rules are well‐defined and consistent	– Require extensive manual effort to develop and maintain
Statistical NLP algorithms	Statistical algorithms use mathematical models and large data sets to understand and process language. These algorithms rely on probabilities and statistical methods to identify patterns and relationships in text data.	+ More flexible and scalable	– Struggle to fully capture context and nuances of language
Hybrid NLP algorithms	Hybrid algorithms combine elements of both symbolic and statistical approaches to leverage the strengths of each. These algorithms use rule‐based methods for certain linguistic tasks and statistical methods for others.	+ Achieve higher accuracy and robustness in NLP applications	– Ensuring smooth interaction between the components is difficult, especially when the rule‐based and statistical components conflict

NLP indicates natural language processing.

Figure 2. AI concepts.

Machine learning and deep learning are subfields within AI. AI indicates artificial intelligence.

Table 3.

AI Algorithm Performance Metrics

Metrics	Definition	Equations
Sensitivity	Also known as true positive rate, quantifies how well a test identifies true positives	TP/(TP + FN)
Specificity	Also known as true negative rate, quantifies how well a test identifies true negatives	TN/(TN + FP)
Positive predictive value	Reflects the proportion of positive results that are true positives	TP/(TP + FP)
Negative predictive value	Reflects the proportion of negative results that are true negatives	TN/(TN + FN)
Area under receiver operating characteristic curve	Evaluates the performance of a binary diagnostic classification method	NA
Precision‐recall curve	Precision and recall are performance metrics that apply to data retrieved from a collection, corpus or sample space	NA
F1‐score	The harmonic means of precision and recall	2TP/(2TP + FP + FN)

AI indicates artificial intelligence; FN, false negative; FP, false positive; NA, not applicable; TN, true negative; and TP, true positive.

AI IN DIAGNOSIS OF HF

AI‐Assisted Echocardiography

Echocardiography is used for a range of diagnostic purposes, from screening to risk stratification. AI applications include advancements in image acquisition, segmentation, and interpretation.

Automated Image Acquisition, View Classification, and Segmentation

Traditional echocardiographic imaging often requires manual adjustments and skillful maneuvering by the operator to obtain high‐quality images from various angles and views. This manual process can be time consuming and prone to variability based on the operator's experience. Echocardiography robots aim to streamline this process by automating the acquisition of echocardiographic images, thereby enhancing efficiency and consistency. Soemantoro et al. introduce a navigation framework for the automated acquisition of echocardiography images that consisted of 3 modules: perception, intelligence, and control. ⁴⁴ Shida et al. proposed a method for finding a parasternal long‐axis view in echocardiography autonomously with a robotic ultrasound system. ⁴⁵ However, for automated image acquisition systems to be applied clinically, limitations such as precise recognition in large variation of anatomy must be considered. Nevertheless, the introduction of an automated echocardiography acquisition system would pave the way for fully end‐to‐end acquisition systems. Additionally, deployment of these platforms may improve point‐of‐care disease diagnosis in primary care and emergency department settings and increase echocardiography availability in resource‐limited locations.

Automated view classification and segmentation are advanced applications of AI in echocardiography. Automated view classification refers to the use of AI algorithms to automatically identify and categorize echocardiographic view. Zhang et al. presented a fully automated echocardiographic interpretation pipeline that includes 23 view classifications. ⁴⁶ The overall accuracy of their model was 84% at an individual image level. However, as authors have pointed out in their research, until the general issue of scalability is solved, their methods may not be suitable for independent use in a clinical setting. Madani et al. trained a convolutional neural network (CNN) to simultaneously classify 15 standard echocardiographic views. ⁴⁷ The model classified among 12 video views with 97.8% overall test accuracy without overfitting. Gearhart et al. developed a computer vision model to autonomously perform view classification on pediatric echocardiographic images. ⁴⁸ The CNN model identified the 27 preselected views with 90.3% accuracy. Zhu et al. developed a deep residual CNN to automatically identify multiple views of contrast and noncontrast echocardiography, including parasternal LV short axis, apical 2‐, 3‐, and 4‐chamber views. ⁴⁹ On the test data set, the overall classification accuracy is 99.1%. One limiting factor for accuracy of view classification is speckle and clutter noise. Kusunose et al. tested 2 types of input methods for the classification of images using DL. ⁵⁰ The best model classified video views with 98.1% overall test accuracy in the independent cohort. The results of these studies provided a foundation for AI‐assisted echocardiographic segmentation and interpretation.

Automated image segmentation involves using AI to partition an identified view into the region of interests, such as left/right ventricle or atrium, ventricular septum, and mitral/tricuspid valves. This process is necessary to quantify certain parameters, such as volume changes and EF. Different AI models have demonstrated their ability to segment 2‐dimensional (2D) and 3‐dimensional echocardiographic images, providing accurate and temporally consistent segmentation maps across the whole cycle. Zhang et al. developed a fully automated pipeline for cardiac chamber segmentation in 5 views. ⁴⁶ U‐Net is a convolutional neural network that was developed for image segmentation. Using 5‐fold cross‐validation on 791 images with manual LV segmentation, their U‐Net‐based model achieved intersection over union metrics of 0.72 to 0.90. However, in their study clearly outliers existed and had large deviations. Leclerc et al. evaluated encoder–decoder DL methods for cardiac structure segmentation. ⁵¹ Their U‐Net model outperformed non‐DL methods. Zamzmi et al. proposed a real‐time system for 3‐dimensional echocardiography analysis. ⁵² The system presented a novel Trilateral Attention Network for cardiac region segmentation. Ta et al. presented a learning network that can simultaneously segment the left ventricle and track its motion. ⁵³ The proposed model can achieve excellent estimates of myocardial segmentation. Wang et al. presented a framework using generative AI technology to produce multiclass RGB masks for image segmentation. ⁵⁴ Their approach outperformed several state‐of‐the‐art models, showcasing improvements in 5 commonly used segmentation metrics.

Automated Image Interpretation, Disease Detection, and Classification

Once automated image segmentation has been performed, AI has been proposed to automate the measurement and assessments. These AI technologies may reduce the time needed for echocardiography technicians and cardiologists to make measurements such as LV wall thickness, ventricular size, atrial size, and LVEF. Additionally, these technologies may yield improvements in intraobserver and interobserver variability. Christensen et al. developed a vision–language foundation model for echocardiography called EchoCLIP. ⁵⁵ It can learn the relationship between cardiac ultrasound images and the interpretations of expert cardiologists across a wide range of patients. The results showed high accuracy in cardiac function assessment and implanted intracardiac devices identification. However, one of the major limitations of this work includes the use of an image encoder instead of a video encoder when echocardiography videos contain important motion‐based information. Ouyang et al. developed a DL algorithm‐EchoNet‐Dynamic using 10 030 echocardiographic videos. ⁵⁶ The accuracy of EchoNet‐Dynamic in estimating LVEF and classifying patients with HF is comparable to that of experienced cardiologists. The AI‐based algorithm incorporated information across multiple cardiac cycles and accurately classified HFrEF (area under the curve [AUC] 0.97). Lau et al. proposed 2 DL‐based echocardiogram interpretation models, DROID‐LA (left atrial) and DROID‐LV, to automate the assessment of standard measurements of LA and LV structure and function. ⁵⁷ The DROID models were trained, tested, and internally validated. In the external validation set, DROID‐LA and DROID‐LV accurately predicted LA and LV linear measures and LVEF. Tromp developed a fully automated DL workflow to classify, segment, and annotate 2D videos and Doppler modalities in echocardiography. ⁵⁸ In the external data sets automated measurements showed good agreement with locally measured values, with a narrow mean absolute error range for LVEF and E/e' ratio, and reliably classified systolic dysfunction and diastolic dysfunction. Tokodi et al. implemented a DL‐based tool to estimate right ventricular EF from 2D echocardiographic videos and benchmarked the tool's performance against human expert reading. ⁵⁹ Using 2D echocardiographic videos alone, the proposed DL‐based tool can accurately assess RV function, with similar diagnostic and prognostic power as 3‐dimensional imaging. Yet, this study is limited by its retrospective design and single‐center database. There are apparent differences in the demographics and clinical characteristics of the internal and external data sets.

Current methods for diagnosing HF rely upon a patient's history, their physical examination, and both laboratory and imaging data. AI‐based methods aim to improve diagnosis through leveraging data found from each of these areas. Sanchez‐Martinez et al. tested the hypothesis that comprehensive ML of LV function at rest and exercise objectively captures differences between HFpEF and healthy subjects. ⁶⁰ ML identified a continuum from health to disease, including a transition zone associated with an uncertain diagnosis. Yet, no external reference was available to validate the ML performance. Tabassian et al. investigated whether timing and amplitude of LV myocardial motion and strain during rest and stress can provide more informative criteria than standard measurements. ⁶¹ They used supervised ML for comparison of the predicted labels with the clinical HF diagnoses. ML of strain rate, compared with standard measurements, gave the greatest improvement in accuracy for the 2‐class task (+23%, P<0.0001), compared with +11% (P<0.0001) using velocity and +4% (P<0.05) using strain. However, no angle corrections were used for strain measurement and therefore could lead to potential biases. Segar et al. used unsupervised clustering analysis to identify distinct phenotypic subgroups in a high‐dimensional, mixed‐data cohort of individuals with HFpEF. ⁶² Chiou et al. established a rapid prescreening tool for HFpEF by using AI techniques to detect abnormal echocardiographic patterns in structure and function on the basis of intrabeat dynamic changes in the LV and the LA. ⁶³ The accuracy, sensitivity, and specificity of the best AI model for detecting HFpEF were 0.91, 0.96, and 0.85, respectively. The model was further validated using an external testing data set, and the accuracy, sensitivity, and specificity became 0.85, 0.79, and 0.89, respectively. Nevertheless, the model performance went from excellent performance to good performance when using an external validation set. Pandey et al. explored a deep neural network model that integrates multidimensional echocardiographic data to identify distinct patient subgroups with HFpEF. ⁶⁴ The model showed higher area under the receiver operating characteristic curve (AUROC) than 2016 American Society of Echocardiography guideline grades for predicting elevated LV filling pressure (0.88 versus 0.67; P=0.01). Chao et al. developed a data‐driven, unsupervised ML approach for diastolic function classification and risk stratification using the LV diastolic function parameters recommended in the 2016 American Society of Echocardiography guidelines. ⁶⁵ Unsupervised ML identified physiologically and prognostically distinct clusters based on 9 diastolic function Doppler variables. Chen et al. developed an AI–assisted system to facilitate the clinical assessment of LV diastolic function. ⁶⁶ AI models successfully achieved LV diastolic function assessment and grading that compared favorably with human experts reading according to guideline‐based algorithms. Moreover, when Doppler variables were missing, AI models could provide assessment by interpreting 2D strain metrics or videos from a single view. However, this method suffers from the disadvantage from being subject to the accuracy of segmentation and strain quantification. Wu et al. explored the potential of natural language processing to improve the detection and diagnosis of HFpEF. ⁶⁷ The study demonstrated that patients with undiagnosed HFpEF are an at‐risk group with high mortality. It is possible to use natural language processing methods to identify likely patients with HFpEF from electronic health record data who would likely then benefit from expert clinical review and complement the use of diagnostic algorithms.

AI‐Assisted ECG

ECG is a cost‐effective, noninvasive diagnostic tool that has endured across clinical medicine over a century after its advent. Efforts to automate ECG interpretation through rule‐based algorithms have been ongoing for decades, because of its reproducible, standardized format. The construction of extensive digital waveform databases has opened the possibility of using DL for HF.

Automated Heart Rate Analysis

Heart rate variability refers to the variation in successive R‐R intervals of the cardiac cycle, reflecting the function of the autonomic nervous system. The relationship between heart rate variability and HF is one of the key research topics in the field of HF. Most of these studies obtain data from public ECG databases and use AI algorithms to distinguish between healthy individuals and patients with HF. The constructed models consistently demonstrate excellent performance. In 2014, Liu and colleagues used support vector machine classifier with 3 nonstandard heart rate variability features to develop a congestive HF classification model that achieved 100% accuracy, sensitivity, and specificity. ⁶⁸ In 2019, a study used the long short‐term memory DL method to identify patients with congestive HF. ECG data from 5 public databases were used for training and testing. ⁶⁹ The existing data were divided into 2 groups based on disease severity for analysis, and 3 different R‐R interval counts (500, 1000, and 2000 samples) were selected for comparison. The analysis results indicated that when using the IDMC‐CHF, NSR, and FD databases, the accuracy of CHF identification across different R‐R intervals was 99.22%, 98.85%, and 98.92%, respectively. When using the NSR‐RR and CHF‐RR databases, the accuracy for different R‐R intervals was 82.51%, 86.68%, and 87.55%, respectively. Although model performance in such studies appears promising, the focus has primarily been on improving ML methods. These algorithms often rely on a large number of heart rate variability parameters, leading to increased model complexity. Additionally, the sample sizes of the selected databases were relatively small, which limits their further application in clinical practice.

Automated ECG Analysis

Cho et al. obtained 39 371 12‐lead ECGs results from 17 127 patients and used a CNN model to detect HFrEF. ⁷⁰ In both internal and external validation cohorts, the AUC for detecting HFrEF was 0.913 and 0.961, respectively. The study provides interpretable features of the model. The lateral and anterior wall leads had a significant impact on the detection outcomes. Additionally, heart rate, QT interval, QRS duration, and T‐axis were highly correlated with the model. Yao et al. conducted a pragmatic clinical trial aimed to assess whether an ECG‐based, AI‐powered clinical decision support tool enables early diagnosis of low EF, a condition that is underdiagnosed but treatable. ⁷¹ The AI algorithm demonstrated excellent performance with a C‐statistic of 0.92 and improved practice when integrated into routine clinical workflows. Yet the limited availability of digitized and well‐labeled ECG data and open‐source data sets may limit development of AI algorithms. Sangha developed a novel CNN‐based approach that can use 12‐lead ECG images for the screening of LV systolic dysfunction. ⁷² The model demonstrated high discrimination power in both internal (AUROC=0.91, area under the precision‐recall curve =0.55) and external validation (AUROC=0.88~0.95, area under the precision‐recall curve =0.45~0.88) for detecting an LVEF <40% at different institutions. Kalmady et al. used XGBoost for simultaneous prediction of 15 different common cardiovascular diagnoses including HF at the population level. ⁷³ The authors employed ResNet‐based DL using ECG tracings and extreme gradient boosting using ECG measurements. DL models outperformed extreme gradient boosting models with about 5% higher AUROC for predicting 8 cardiovascular conditions. ECG‐based prediction models demonstrated good‐to‐excellent prediction performance in diagnosing HF. A DL algorithm for ECG‐based HF identification was developed and validated for early diagnosis of HF. ⁷⁴ The algorithm was superior in detecting HFrEF compared with logistic regression and random forest ML algorithms.

AI‐Assisted CMR

AI is poised to transform the field of cardiovascular magnetic resonance (CMR) by addressing its traditional limitations such as lengthy exam times, high costs, and the need for expert manual analysis. By automating complex image processing and enhancing diagnostic accuracy, AI can significantly improve the efficiency and accessibility of CMR in the assessment of HF.

Multiple vendors implement CNN‐based denoising for image reconstruction, which can result in higher‐quality image reconstruction with shorter acquisition times. Kucukseymen et al. developed a supervised ML model to predict HF hospitalization in patients with HFpEF using noncontrast CMR imaging. ⁷⁵ The study compared a basic clinical model with an advanced model using the XGBoost algorithm, demonstrating that the machine learning model significantly improved prediction accuracy (AUC: 0.81 versus 0.64). However, this study was limited by its retrospective nature and relatively small sample size. Xie et al. developed the 4‐dimensional self‐supervised learning framework of HF classification, which uses self‐supervised learning to enhance HF classification from cine CMR imaging by integrating 3‐dimensional spatial and temporal information. ⁷⁶ This advanced approach achieved an AUC of 0.8794 and an accuracy of 0.8402, demonstrating its potential to improve diagnostic accuracy and support personalized treatment decisions. Although this method has outstanding performance, more data could have been collected to further improve the classification performance. Furthermore, due to graphics processing unit memory limitation, the weights of the Siamese network were fixed during training. A combined training of 2 networks could better integrate the information of all dimensions. Lehmann et al. proposed an AI‐enhanced CMR imaging method for diagnosis of cardiovascular disease classification and diastolic filling pressure. ⁷⁷ A total of 6936 patients were analyzed, and 4390 were included in the final cohort. AI models demonstrated strong accuracy in predicting various parameters related to heart disease. The AI models could help classify diseases and predict LV end‐diastolic pressure, adding value to CMR imaging. The study highlights the potential of AI‐assisted CMR to improve noninvasive cardiac assessments, offering practical applications for evaluating cardiac function and diagnosing HF. Wang et al. developed and validated computerized CMR interpretation for screening and diagnosis of 11 types of cardiovascular diseases in 9719 patients. ⁷⁸ The screening and diagnostic models achieved high performance (AUC 0.988±0.3% and 0.991±0.0%, respectively) in both internal and external data sets. This proof‐of‐concept study holds the potential to substantially advance the efficiency and scalability of CMR interpretation. Radhakrishnan et al. developed a cross‐modal autoencoder framework by unsupervised ML algorithm to integrate myocardial structural information from CMR and myoelectric information from ECG for holistic representations of cardiovascular state. ⁷⁹

AI‐Assisted Coronary Angiography

A major cause for end‐stage HF is coronary artery disease. The gold standard to diagnose coronary artery disease is coronary angiography. In the field of coronary angiography, AI has shown potential in assisting with image acquisition, interpretation, and risk stratification.

Avram et al. used neural networks to develop a fully automated coronary angiography interpretation and stenosis estimation system to interpret angiographic coronary artery stenosis. ⁸⁰ The coronary angiography interpretation and stenosis estimation system is a pipeline of multiple deep neural network algorithms. The training set used 13 843 angiographic studies. Algorithms were internally and externally validated, with positive predictive value, sensitivity, and F1 score achieving >90% for projection angle identification and 93% for left/right coronary artery angiogram detection.The coronary angiography interpretation and stenosis estimation system exhibits an AUC of 0.86 for obstructive coronary artery disease stenosis prediction. Yet, this work was limited by 1 notable drawback. The authors used training labels derived from physician visual estimation and clinically generated stenosis values. Inherent variability still exists, which could place an artificial ceiling on observed algorithmic performance. Du et al. successfully developed an automated, multimodal system for analyzing and quantifying coronary angiography, incorporating various elements such as the detection of coronary artery segments and the identification of lesion morphology. ⁸¹ A data set of 20 612 angiograms was retrospectively gathered, with 13 373 of them labeled for coronary artery segments and 7239 labeled for specific lesion morphology. For segment prediction, the accuracy rate reached 98.4%, with a sensitivity of 85.2%. Naghavi et al. investigated the performance of AI‐enabled cardiac chambers volumetry in coronary artery calcium scans for prediction of incident HF. ⁸² AI‐enabled cardiac chambers volumetry in coronary artery calcium scans significantly outperformed NT‐proBNP and the Agatston score in predicting incident HF over 15 years, with a higher time‐dependent AUC (0.86 versus 0.74 and 0.71, respectively). Adding AI‐coronary artery calcium to clinical risk factors also substantially improved risk prediction, as demonstrated by a significant increase in the net reclassification index for HF prediction.

Wearable Devices

Wearable devices have the potential to enable widespread AI‐driven screening. Several models have recently been validated using wrist‐worn wearables. ECGs from wearable smartwatches, collected outside of clinical settings, can effectively identify patients with cardiac dysfunction, which is often life threatening and may be asymptomatic.

Khunte reported a novel strategy that automates the detection of hidden cardiovascular diseases such as LV systolic dysfunction, designed for noisy single‐lead ECGs from wearable and portable devices. ⁸³ A total of 385 601 ECGs were used to develop both a standard model and a noise‐adapted model. Both models showed similar performance on standard ECGs, achieving an AUROC of 0.90 for detection of LVEF <40%. Despite apparent merits, this study has 1 limitation that warrant consideration. The model was developed among patients with both ECG and echocardiography. Because the training population had a clinical indication for echocardiography, selection bias exists. This limits the broader use of the algorithm for screening tests for LV systolic dysfunction among those without any clinical disease in the real world. Attia et al. conducted a prospective analysis where Mayo Clinic patients were invited via email to download a Mayo Clinic iPhone app that transmits watch ECGs to a secure data platform. ⁸⁴ In this study, 2454 patients were digitally enrolled, and 125 610 ECGs were submitted. The AI algorithm identified patients with low EF (defined as ≤40%) with an AUROC of 0.885 and 0.881, using the mean prediction within a 30‐day window or the closest ECG relative to the echocardiogram that determined the EF, respectively. Inan et al. used wearable devices to measure electrical and mechanical aspects of cardiac function for HF state assessment. They devised a method based on comparison of the similarity of the structure of seismocardiogram signals after exercise and at rest and reached the conclusion that ML‐assisted wearable technologies recording cardiac function can assess compensated and decompensated HF states.

Limitations of AI

To date, the application of AI in cardiovascular diseases have been promising. However, as mentioned previously, limitations exist regarding both AI technology itself and infrastructures in the medical setting that hinder implementation of AI in daily clinical practice.

Data Diversity and Availability

First and foremost, large medical data sets are rarely available due to privacy concerns. However, for the robust development of AI diagnostic models, access to big data is essential. Additionally, in the real‐world hospital setting, different medical data are often stored across multiple servers and, in some cases, in paper records. Even if AI produces highly accurate prognostic models, their effectiveness could be limited if hundreds of parameters for prediction are scattered across various systems and need to be entered manually.

Biases in Algorithm Development

Second, biases in AI algorithms often arise from nonrepresentative data sets, leading to biased predictions or outcomes in new populations. Overfitting is another common issue that causes poor generalizability of AI models. An overfit model performs well on the training data but poorly on validation or test data sets. Several studies demonstrated poor generalizability of HF scoring systems in new populations. For instance, the famous Framingham Risk Score could overestimate (or underestimate) risk in populations other than the US population. ⁸⁵ , ⁸⁶

Interpretability in Clinical Context

The interpretability of AI models is often questioned in the clinical context due to their “black box” nature. A key limitation of current AI approaches is their inability to establish causal inference. Additionally, inaccurate analyses and insufficient reporting hinder reliable assessments and may lead to misleading interpretations. Therefore, AI‐generated results should be carefully interpreted within the framework of medical knowledge.

Future Direction

Despite these limitations, AI has shown early promise in the diagnosis of HF. Future directions for AI‐assisted echocardiography will probably focus on using guidance tools for image acquisition, increasing image interpretation efficiency and reliability, and automating disease detection. The future goal for AI‐assisted CMR might be to further improve the speed of image analysis. Real‐time telemetry is another bright direction for AI‐assisted HF diagnosis.

CONCLUSIONS

In summary, the integration of AI into the diagnosis of HF represents a transformative advancement in cardiovascular medicine. The use of AI technologies, including ML algorithms, DL models, and predictive analytics, has shown significant promise in enhancing the accuracy, efficiency, and timeliness of HF diagnoses. By leveraging large data sets and sophisticated computational techniques, AI systems can identify patterns and correlations that may be overlooked by traditional diagnostic methods, leading to earlier detection and personalized treatment strategies.

Despite the impressive progress, there remain several challenges and limitations that must be addressed, including the need for high‐quality, diverse data sets; the potential for algorithmic biases; and the requirement for clinical validation to ensure real‐world applicability. As research and technology continue to evolve, the role of AI in HF diagnosis is likely to expand, offering new opportunities for improving patient outcomes and advancing the field of cardiology.

Sources of Funding

This work was supported by grants from the National Natural Science Foundation of China (grant number 82230066 & 82402 304) and the Fundamental Research Funds for the Central Universities (grant number YCJJ20241409).

Disclosures

None.