The Emerging Role of Artificial Intelligence in Valvular Heart Disease

Abstract

Valvular heart disease (VHD) is a morbid condition in which timely identification and evidence-based treatments can lead to improved outcomes. Artificial intelligence broadly refers to the ability for computers to perform tasks and problem solve like the human mind. Studies applying AI to VHD have used a variety of structured (eg, sociodemographic, clinical) and unstructured (eg, electrocardiogram, phonocardiogram, and echocardiograms) and machine learning modeling approaches. Additional researches in diverse populations, including prospective clinical trials, are needed to evaluate the effectiveness and value of AI-enabled medical technologies in clinical care for patients with VHD.

Introduction

Valvular heart disease (VHD) has been found in 14% of patients referred for echocardiographic workup of suspected heart failure (HF)¹ and 21% of patients hospitalized for HF.² The epidemiology of VHD varies around the world. In high-income countries, VHD is predominantly due to degenerative disease, whereas in low-income and middle-income countries, VHD is predominantly due to rheumatic heart disease.³ Even with this variance, the global incidence of VHD has increased by 45% in the last 30 years due to an aging population,⁴ contributing to burden of degenerative disease.^{4, 5} Consequently, the proportion of HF secondary to VHD is likely to increase ³ Aortic and mitral valvulopathies predominate in prevalence, particularly calcific aortic valve disease and degenerative mitral valve disease.^{1, 3} Substantial advances in the management of severe VHD have been made in recent decades with the development of transcatheter approaches to valve repair and replacement. Transcatheter aortic and mitral valve procedures have increased by nearly 5-fold and 10-fold, respectively, in recent years with growing favorable outcomes.^{6, 7} Tricuspid transcatheter valve repair and replacement are also emerging therapies and are currently under investigation.^{8, 9}

VHD is a ripe area for the development and testing of artificial intelligence (AI) or machine learning (ML) approaches. A systematic review of AI/ML in cardiovascular medicine found that less than 3% of studies were applied in this disease area.¹⁰ Cardiovascular medicine is increasingly digital, with widespread integration of electronic health records (EHRs) in clinical practice, repositories of cardiovascular diagnostic data including echocardiograms, electrocardiogram (ECG), phonocardiograms (PCGs), and cardiac computed tomography (CT) and MRI. The vast amount of data coupled with the increased availability of high-performance computing will be conducive to the continued growth in AI and ML research and application development to improve detection, risk prediction, and management of VHD.

AI broadly refers to the ability for computers to perform tasks and problem solve like the human mind. ML is a subset of AI that emerged from the fields of computer science and statistics. ML relies on input, specifically large collections of data, to recognize patterns, entities, and the connections between them. The discipline of ML explores the analysis and construction of algorithms that can learn from and make predictions based on data. ML algorithms build a mathematical model based on sample data, known as training data, to make predictions or decisions without being explicitly programmed to do so.¹¹ An increasingly utilized subset of ML is deep learning (DL), or artificial neural networks (ANNs). This specifically refers to algorithms that are modeled after the human brain containing artificial “neurons” arranged in successive layers culminating in a decision function. This relationship is illustrated in Fig. 1.

Fig. 1.

Artificial intelligence and machine learning. Relationship between artificial intelligence, machine learning, and “deep” learning. CNN, convolutional neural network; DNN, deep neural network; GAN, generative adversarial network; KNN, k-nearest neighbor; PCA, principal components analysis; RF, random forests; RNN, recurrent neural network; SVM, Support vector machines; VAE, variational autoencoder. (From Wehbe RM, Khan SS, Shah SJ, Ahmad FS. Predicting High-Risk Patients and High-Risk Outcomes in Heart Failure. Heart Fail Clin. 2020;16(4):387–407.)

Typically, an ML dataset will be split into 3 categories: training, validation, and testing. The training subset is used to fit an ML model. The validation subset is to test the model and tweak certain hyperparameters because it is trained to optimize performance. Finally, the test subset is used to evaluate the fit, or model performance of the ML model. It is important that the test subset be segregated from the rest of the data so it cannot be included in the training and validation.

Two main approaches, supervised and unsupervised learning, have been widely applied to a variety of clinical problems and tasks in biomedical science. Supervised learning requires human input to label input and output training data. A model will learn the relationship between the inputs and outputs, and will adjust to make the correct answer. Therefore, supervised models are generally used for classification or regression (predicting outcomes). Unsupervised learning identifies intrinsic patterns within data without human intervention, and the output of the model is generally patterns or trends. These models generally include clustering, association tasks, dimensionality reduction, or factorization. Unsupervised learning can be used in exploratory data analysis to inform supervised learning tasks. Other types of learning frameworks, such as reinforcement learning, have been applied in health care. Several reviews have described in detail different ML approaches, applications, strengths, and limitations for addressing problems in health care more broadly and within cardiovascular medicine.^{10, 12–17}

In this review, we highlight areas AI/ML methods and their application may address the following unmet needs in the care of patients with VHD: (1) screening for VHD, (2) assessment of severity of disease, (3) risk prediction, and (4) phenomapping ( Fig. 2 ). Later, we discuss some of the challenges and future directions of AI/ML in VHD.

Fig. 2.

Clinical applications of AI/ML in valvular heart disease. Four main tasks for AI/ML in valvular heart disease are reviewed in this article: (1) screening to valvular heart disease, (2) assessment of severity of disease, (3) risk prediction, and (4) phenomapping. For each task, a variety of data types and models have been applied, predominantly in retrospective datasets. Each of these has the potential to lead to improvements in clinical care in 4 main areas but additional, prospective clinical studies are needed.

Screening for valvular heart disease

Several studies have shown that VHD remains underdiagnosed.^18–22 In the United States alone, national prevalence of VHD affects 2.5% of the population.¹⁹ Most patients are not screened for VHD until they report the onset of symptoms even though most cases remain asymptomatic for many years.^{19, 22} This presents an opportunity for early detection before the lesion becomes more severe, causes symptoms, or leads to ventricular dilation or dysfunction. More studies are calling for early identification and treatment of VHD,^{23, 24} and thus timely identification has implications for disease management and improvement of patient outcomes.^25–27 The gold standard of detecting VHD is via echocardiogram¹⁸ but due to cost and resource limitations it is usually reserved for either symptomatic patients or those for whom there is high clinical suspicion of VHD based on physical examination. Other cardiac examinations that are more widely indicated can help recognize patients with asymptomatic VHD that should be referred for an echo. Below we explore examples of studies that are using ML strategies to help identify patients with VHD through auscultation and ECG. Although these examples show that these technologies can identify patients with underdiagnosed VHD, prospective randomized controlled trials and cost-effectiveness analyses are needed to inform recommendations from guideline-writing organizations, such as various cardiovascular societies and the US Preventive Services Task Force.

Auscultation

The stethoscope was invented in 1818 by René Laennec. It has remained the most inexpensive and accessible modality to detect VHD as heart sounds can reveal the cause of a lesion as well as the severity. Although every clinician learns to auscultate as a foundational part of the physical examination, the ability to detect and interpret murmurs varies widely.^{28, 29} Thus, researchers have explored ways to aid clinicians in this task.

Attempts to classify pathologic condition in PCGs, or visual plot representation of heart sounds, has been performed for more than 50 years. This was accelerated by the 2016 PhysioNet/Computing in Cardiology Challenge,³⁰ which published an open-access database of 3126 heart sound recordings. Researchers were tasked with identifying PCGs as normal, abnormal, or poor signal quality. The 3 top performing entrants used ensemble methods, or combination of different classifiers to improve accuracy. However, each team used different classifiers, demonstrating that different approaches could have comparable performance. ³¹ The feature extraction stage, which transforms the data into derived values, or features, to reduce redundancy was deemed key to effective models.³¹

Digital stethoscopes have enabled the recording and automated classification of heart sounds. These stethoscopes work by converting acoustic sound to electronic signals, which can be amplified for the examiner.³² Many models can transmit the information to a mobile app or computer in the form of an audio file or PCG.³²

One AI-enabled digital stethoscope that detects and characterizes murmurs in adults and pediatric patients has received Food and Drug Administration (FDA) clearance and Conformite Europeenne (CE) mark. This was supported by a study from Chorba and colleagues,³³ who created and tested a deep convolutional neural network (CNN) model that classifies PCGs in 3 categories: heart murmur, no murmur, or inadequate signal. To determine the accuracy of the signal quality and murmur detection performance, the algorithm output was compared with annotations from 3 cardiologists. Algorithm performance had a sensitivity and specificity for detecting murmurs of 76.3% and 91.4%. The study also used the gold standard echocardiogram to detect “clinically significant” (more than moderate) aortic stenosis (AS) and mitral regurgitation (MR). Detection of AS at the aortic position was 93.2% sensitive and 86.0% specific. The detection of MR at the mitral position was 66.2% sensitive and 94.6% specific.

Electrocardiogram

Besides auscultation, ECG is another relatively inexpensive and ubiquitous noninvasive test to evaluate the heart. In current clinical practice, valvular diseases are not directly diagnosed on ECG but evidence of structural heart disease, such as left ventricular hypertrophy, can provide clues for VHD leading to left ventricular remodeling. Several studies have examined whether DL can be applied to ECGs to detect VHD in patients.

The ability to detect and predict the future development of severe AS is of interest because the condition is associated with a mortality rate of 40% to 50% within 1 year without valve replacement.^{34, 35} Multiple groups of investigators have trained a specific neural network called a CNN, which are often used for image classification and process an image in a grid-like fashion, using ECG-Echo pairs. Kwon and colleagues³⁶ used 39,371 ECGs from patients at a cardiology teaching hospital to train a CNN to detect AS, and then performed external validation on the algorithm with 10,865 ECGs from patients at a community hospital, which showed generalizability with an area under the curve (AUC) of 0.861, sensitivity 0.80, and specificity 0.78. They also used a saliency map, which focused on the T wave, to highlight the parts of the ECG that were used to predict presence of AS. This type of technique has been used to increase confidence for clinicians to accept the prediction results because the user sees that the algorithm is using an important feature to make a prediction rather than some artifact in the background. Results from Cohen Shelley and colleagues³⁷ reinforce these findings. The authors trained a CNN to identify moderate-to-severe AS in 258,607 adults with ECG-TTE pairs, with a comparable AUC of 0.85, sensitivity of 0.78, and specificity of 0.74 when using just ECG morphology alone. Their saliency map similarly highlighted the TP segment as an area of importance. However, several studies have raised concerns about the accuracy and reliability of saliency maps for DL algorithms and emphasize the need for prospective testing of algorithms to generate the evidence base to support adoption over the reliance of saliency maps and other techniques to increase confidence of end-users.^{38, 39}

Because the overall incidence of VHD is relatively low, the rECHOmmend study⁴⁰ combined moderate-to-severe valvular disease (AS, AR, MS, MR, tricuspid regurgitation [TR]) with structural changes of reduced EF less than 50%, or ventricular septal thickness greater than 15 mm to achieve increased positive predictive value (PPV) for their models. Authors marked ECGs as positive for a given condition if it was acquired within a year before the patient’s first positive echocardiogram for the label. Then they trained a CNN to predict presence or absence of any of the 7 labels with ECG tracings and a range of input data including demographics, laboratories, findings, and ECG measurements. Although the model performed best with all inputs, the area under the receiver operating characteristic (AUROC) remained high (0.91) with age, sex, and ECG tracing data. This potentially enables more targeted referral for echocardiography to help detect underdiagnosed VHD, although prospective testing is needed to determine the clinical utility and cost-effectiveness of these approaches.

Assessment of severity for valvular heart disease

The studies that explore severity predominantly use echocardiography as an input, as echocardiography remains the primary modality for diagnosis and evaluation of VHD.¹⁸ Echocardiograms are one of the least expensive and most accessible forms of cardiac imaging.^41–43 Between 2001 and 2011, there were more than 7 million echocardiograms performed annually in the United States, with its use increasing by more than 3% yearly.⁴¹ AI has been used for a variety of tasks related to echocardiography,^44–47 including view acquisition, view classification,⁴⁸ quality assessment,⁴⁹ automation of measurements and observations,^{50, 51} and disease identification, such as cardiac amyloidosis⁵² and hypertrophic cardiomyopathy.^53–55

The use of AI to diagnose VHDs via echocardiographic images represents an emerging area of research. For example, Yang and colleagues⁵⁶ developed a 3-stage DL framework that evaluated view classification, screening for VHDs (aortic and mitral lesions), and when positive, performed segmentation to provide metrics related to severity. The investigators used 2 different approaches to detect regurgitant and stenotic lesions. A CNN was trained to detect color in the left ventricular outflow tract (LVOT) or left atrium for AR and MR, respectively. It also measured anatomic structures to localize the regurgitation in time and space. For stenotic lesions, the algorithm identified the PLAX-2D and A4C views as input for the diagnosis and then calculated valve area with Doppler to calculate mitral stenosis (MS), and the Vmax and mean pressure gradients for AS. Overall, the DL model provided varying agreement when compared with measurements by 2 physician echocardiographers with Intersection of Union (IoU) metrics. IoU is also referred to as the Jaccard Index, is a common metric used in image segmentation. The lowest IoU was for AR jet 0.63, whereas the highest IoU was for the left atrial area 0.86. Next, the model achieved high performance for identifying valvular stenosis and regurgitation with AUC for MS 0.95, MR 0.97, AS 0.91, and AR 0.97 using retrospective data. Finally, the study concluded that the accuracy of the degree of valve lesion severity was within the bounds of normal practice when compared with 2 experienced physicians. They also validated their results in prospective, consecutive cases (n=1374) to demonstrate generalizability and adequate performance outside of the dataset used for training. This study describes in detail the development and testing of a pipeline for a complex set of tasks and posted their code in a public repository, both of which are essential to promoting reproducibility and advances in this field.

In a study by Sengupta and colleagues,⁵⁷ used both unsupervised and supervised ML classifiers to predict high-severity or low-severity AS. Their model was trained on echocardiographic data from 1052 patients with AS and derived high-severity and low-severity groups of AS, they validated their association of severity with other markers of AS such as AV calcium score on CT, markers of myocardial damage on CMR, and biomarkers such as brain natriuretic peptide (BNP). The ML model was able to correctly classify 99% of patients with definitive echocardiographic features of severe AS into the high-severity group. Notably, it was also able to reclassify the remaining patients with either nonsevere AS or inconclusive discordant echocardiographic findings without additional tests. When compared with conventional strategies for AS severity classification, the ML severity classification improved discrimination (integrated discrimination improvement: 0.07; 95% CI: 0.02–0.12) and reclassification (net reclassification improvement: 0.17; 95% CI: 0.11–0.23) for the outcome of aortic valve replacement (AVR) at 5 years. The investigators show that ML can analyze echocardiographic measurements to accurately classify AS severity in most patients with potential to optimize timing of AVR.

Recent preliminary results from a commercial AI-echo company sought to detect and characterize aortic stenosis purely from B-mode images without spectral or color Doppler. Trained on 30,000 echoes (parasternal long axis and short axis and apical 5-chamber views), a test set demonstrated detection of moderate-or-severe AS with sensitivity of 91% and specificity of 94%.⁵⁸ Another commercial AI company developed the ability to identify severe AS with clinical characteristics and echocardiogram measurements without LVOT data. The model performed equally independent of AS gradients.⁵⁹ An increasing number of companies are developing technology in this space and will be subject to FDA review.

In addition to classifying VHD severity, AI can be used to acquisition of images by personnel without earlier ultrasound experience to obtain diagnostic quality echoes.⁶⁰ With further development and validation for the destination of VHD, this use case of AI could be scaled to resource-limited settings without access to trained sonographers to better detect conditions such as rheumatic heart disease.

Besides using echocardiogram images and videos, natural language processing (NLP) has been used to extract valve severity from echocardiogram reports given the challenges with capturing and using structured data generated by echocardiographers and the limited accuracy of International Classification of Diseases (ICD) coding.^{61, 62} NLP can be rule-based or use ML. Solomon and colleagues⁶¹ developed an NLP algorithm with the goal of making sure that all patients with VHD are appropriately followed. The NLP algorithm had a 99% PPV and 99% NPV for identifying prevalent AS. They found that among those classified as having AS by the NLP, only 64.6% of patients had an ICD diagnosis code of AS. They also found that the NLP-derived hemodynamic parameters were concordant with the physician designated levels of severity. Identifying and recording a patient’s diagnosis in the medical record and developing care pathways can help increase rates of appropriate follow-up in accordance to clinical guidelines.¹⁸

Another strategy to make sure a diagnosis is not missed is through clinical decision support (CDS). CDS are tools embedded in a clinical workflow, such as the EHR, to help improve decision-making. A multicenter study of 35 institutions used 1,147,157 echocardiograms to build a tool to help identify patients who may mistakenly not have received a diagnosis of severe AS called the Diagnostic Precision Algorithm.⁶³ The tool uses echocardiographic parameters to determine if patients with similar measurements would be likely to have severe AS, then it prioritizes patients for review based on how likely these valve measurements would be considered severe. The NLP is used to detect that the echocardiogram report was not categorized as severe AS. Overall, the Diagnostic Precision Algorithm was able to predict the proportion of severe AS versus the actual proportion with an average error of 2.1% in the prediction model with left ventricular ejection fraction (LVEF) and an error of 2.2% without LVEF. As the authors highlight, the algorithm demonstrated high accuracy and yields good potential for application to assist physicians to follow-up with patients with potentially undiagnosed severe AS.¹⁸

Alternatively to clinician CDS, similar to HF, workflows could be created to either support a population health approach in which a member of the clinical team reviews cases and works with the patients’ primary team to ensure adherence to guideline recommends or directly nudge the patient to discuss evidence-based care for VHD with their clinician.^{64, 65}

Risk prediction

Risk prediction models are equations that use patient risk factors to estimate the probability of a health-care outcome. Well-known risk-prediction models include the Well’s Criteria⁶⁶ for pulmonary embolism or the Quick SOFA (qSOFA) score⁶⁷ for identifying high-risk mortality from sepsis in critically ill patients. ML is just one tool among many that can be used to create these models. In VHD, ML has been applied largely in the area of predicting outcomes and response to AVR rather than determining the risk of developing VHD.

The discrepancy of the aortic valve area and low transvalvular gradient makes the management of patients with low gradient AS challenging.⁶⁸ Namasivayam and colleagues⁶⁹ developed the Aortic Stenosis Risk (ASteRisk) with the goal of predicting outcomes within 5 years of follow-up for patients with moderate-to-severe AS that would be interpretable and also effective in patients with low-gradient AS. Using bootstrap Lasso regression (Least Absolute Shrinkage and Selection Operator), they selected a subset of variables for a logistic regression model that minimized prediction error (AUC = 0.74). They limited the ASteRisk score to include 9 readily available variables that include a mix of echocardiogram measurements and clinical history to maximize the ease of use in practice. Some of the risk factors included aortic valve area, mean gradient, energy loss, HF, hyperlipidemia, and CKD. They found the ASteRisk model superior to predicting all-cause mortality or AVR in moderate-to-severe AS patients for years 2 to 5, versus a baseline model that used more conventional risk factors (mean transvalvular gradient, aortic valve area, age, and LVEF). The model also worked in patients with low-gradient AS. They tested generalizability on patients from another institution that identified patients at high risk of death at 1 to 5 years. Additional external validation and, ideally, prospective testing, will help elucidate the clinical usefulness of this model for clinical decision-making.

CMR can provide information about not only the structure and function of the heart but also the degree of myocardial fibrosis.⁷⁰ Kwak and colleagues set out to investigate if how markers of damage on CMR could predict prognosis in severe AS and help guide timing for AVR.⁷¹ Investigators used a supervised ML method called Random Forest, a technique that outputs the average result of multiple decision trees, to determine the predictors of death in patients with severe AS undergoing AVR. The study used clinical, echocardiogram, and CMR variables but remained focused on 4 CMR variables, which emerged as most predictive of all cause mortality: extracellular volume fraction (ECV%), late gadolinium enhancement (LGE%), right ventricular ejection fraction (RVEF), and indexed left ventricular end-diastolic volume (LVEDVi). They analyzed the parameters of each variable associated with markedly worse prognosis and built an AS-CMR risk score, in which a point was given for ECV% greater than 27%, REVF 50 or lesser or greater than 80%, LGE% greater than 2%, and LVEDVi 55 or lesser or greater than 80 mL/m² . Noting the importance of RVEF as an important prognostic marker, authors posited that these patients may benefit from TAVR, where RV function is generally maintained, versus SAVR, which is associated with RV dysfunction after surgery.

Phenomapping

Many have posited that ML will be a critical tool in achieving personalized medicine, where medical decisions are tailored to an individual’s phenotype. VHDs, even of the same severity, are heterogenous syndromes similar to other cardiovascular conditions. ML methods have been used in multiple studies to define phenogroups, or groups of patients with similar clinical characteristics, and differential risk of adverse outcomes.

In aortic stenosis, valve area, transaortic gradient, and maximum aortic velocity are the primary quantitative measurements to stratify severity. One study of patients with severe aortic stenosis used 3 unsupervised, independent methods to identify 4 phenogroups.⁷² Then, the investigators used a supervised algorithm to predict survival outcomes by each cluster. Each group had an increased amount of myocardial damage and mortality, regardless of the severity of AS. Therefore, authors concluded cardiac remodeling and progression of systolic and diastolic dysfunction plays a large role in prognosis. Guidelines recommend patients with symptomatic, severe AS undergoing evaluation for aortic valve replacement¹⁸ but these results suggest cardiac remodeling may be another factor during the evaluation for timing of intervention in asymptomatic patients.⁷²

Several other studies of AS have used unsupervised methods to identify clusters with different rates of adverse events after valve replacement. Lachmann and colleagues⁷³ used unsupervised clustering among patients undergoing TAVR for severe AS, identifying 4 clusters with variations of pulmonary hypertension, LVEF, and prevalence of other valvular defects. Then, investigators tested the ability to reproduce the phenogroups prospectively by categorizing the patients into clusters with an ANN, following their clinical course, and measuring the different rates of adverse events between groups. Kwak and colleagues⁷⁴ grouped 398 patients with moderate-to-severe AS into 3 clusters that were broadly defined as patients with (1) cardiac dysfunction, (2) comorbidities (particularly ESRD), and (3) healthy AS with neither cardiac dysfunction nor comorbidities. Although AS severity between clusters did not differ, mortality rate was significantly different between clusters. During a median 2.4-year follow-up, the comorbidities cluster had the highest mortality rate of 19.8% versus the healthy AS cluster had the lowest mortality rate of 6.0% during the 2.4 years.

Phenomapping has also produced valuable insights about MR. There are many causes of secondary MR.⁷⁵ Bartko and colleagues⁷⁶ focused on phenomapping in sMR in patients with HF with reduced ejection fraction (HFrEF) using echocardiogram measurements and clinical variables. Principal component analysis, which is an unsupervised learning technique for dimensionality reduction, identifies the most informative variables in the data. Investigators found 4 phenogroups with different morphologies of varying LV and LA size and dilation. Each cluster had stepwise progression in sMR severity but regurgitant fraction peaked in Cluster 3, which was characterized by small LV and extensive LA dilation. The approach suggested that clinicians should consider not only LV remodeling but also LA and mitral valve annulus. Notably, authors published a simplified calculator that clinicians can use to identify which cluster an individual patient belongs to. Further translating these phenogroups into practice, the investigators remarked that these phenotypes may help explain difference in outcome from 2 landmark MitraClip (Abbot Laboratories) clinical trials, and the need to better understand patient selection and MR phenogroups.^{77, 78}

The 2020 ACC/AHA Guidelines for Management of Patients with Valvular Heart Disease noted thresholds for interventions are expanding because of randomized clinical trials.¹⁸ Whether phenogroup mapping will provide clinically meaningful information and ultimately inform the decision-making process for valvular interventions remain unanswered questions.

Discussion

We have discussed some of the unmet needs in the early detection, diagnosis, and management of VHD that may be addressed by AI-enabled technologies. Examples include screening for VHD with ECG or PCG, improving the classification of the severity of specific types of VHD, and phenomapping and risk prediction to identify subgroups at differential risk for worsening disease and cardiovascular events ( Tables 1 and 2 , see Fig. 2 ). Most research currently leverages data from a single modality or multimodal tabular data but a growing area AI and VHD research in the coming years will be developing models leveraging a combination of unstructured and structured, multimodal, longitudinal data.⁷⁹ Research will also likely increase training models using self-supervised learning, which has outperformed other DL algorithms for several language tasks and now is being more widely tested in computer vision, signal analytics, and other data types.¹⁴

Table 1.

Artificial intelligence applications included in this review

Valve	Valvulopathy	Screening for VHD	Severity of Disease	Risk Stratification	Phenomapping
Aortic	Stenosis Regurgitation	x x	x x	x	x
Mitral	Stenosis Regurgitation	x x	x x		x
Tricuspid	Stenosis Regurgitation	x	Not present
Pulmonary	Stenosis Regurgitation		Not present

Table 2.

Highlighted artificial intelligence/machine learning algorithms in valvular heart disease

Category	Title		Authors	ML Approach			Model Training Data	Objective		Key Findings
Screening for VHD	rECHOmmend: An ECG-Based Machine Learning Approach for Identifying Patients at Increased Risk of Undiagnosed Structural Heart Disease Detectable by Echocardiography		Ulloa-Cerna AE, Jing L, Pfeifer JM, et al.	Convolutional Neural Network with XGBoost			2,925,925 ECGs from 631,710 patients, clinical demographics (age, sex), laboratories, vitals	Detect AS, AR, MS, MR, TR, reduced EF <50% or ventricular septum thickness >15 mm		Model using just ECG tracing, age, sex had AUC = 0.91 at sensitivity 0.90, with a PPV of 0.42 for clinically meaningful disease. Validated at 10 other sites, with AUCs 0.79–0.93
Severity of Disease	Automated Analysis of Doppler Echocardiographic Videos as a Screening Tool for Valvular Heart Diseases		Yang Feifei, Chen Xiaotian, Lin Xixiang, et al.	Convolutional Neural Network			1,335 echocardiogram Doppler videos and images	Echocardiogram view classification, segmentation, identification of stenosis vs regurgitation, and grading of severity		In retrospective dataset for identifying aortic and mitral valve stenosis and regurgitation all accuracy ≥0.87, sensitivities ≥0.82, specificities ≥0.88. For prospective cohort, model performance similar for stenosis, and slightly decreased for regurgitation
Risk Stratification	Markers of Myocardial Damage Predict Mortality in Patients With Aortic Stenosis		Kwak S, Everett RJ, Treibel TA, et al.	Random Survival Forest			CMR measurements, echocardiogram measurements, demographics, clinical variables from 440 patients	Identify markers of myocardial fibrosis in CMR that are prognostic in AS for patients undergoing AVR		Created the AS-CMR score to identify patients at high risk of mortality post-AVR. Patient gets one point for ECV % >27%, RVEF ≤50 or >80%, LGE % >2%, LVEDVi ≤55 or >80 mL/m2. Each point has statistically significant difference in 3 y mortality
Phenomapping		Principal Morphomic and Functional Components of Secondary Mitral Regurgitation	Bartko PE, Heitzinger G, Spinka G, et al.		Principal Component Analysis	Morphology and functional parameters from echocardiograms (32 variables) from 383 patients			Identify different phenogroups of secondary MR patients with HFrEF	Identified four clusters of patients with atrial and ventricular morphologies, and regurgitant fractions. Each cluster had statistically significant survival times. Published an online tool to calculate for individual patient

Abbreviations: AR, aortic regurgitation; AS, aortic stenosis; AVR, aortic valve replacement; CMR, cardiovascular magnetic resonance; ECG, electrocardiogram; ECV%, extracellular volume fraction; EF, ejection fraction; HfrEF, heart failure with reduced ejection fraction; LGE%, late gadolinium enhancement; LVEDVi, indexed left ventricular end-diastolic volume; MR, mitral regurgitation; MS, mitral stenosis; RVEF, right ventricular ejection fraction; TR, tricuspid regurgitation.

Another expanding area of research will be the application of large language models. ChatGPT and BioGPT, released by OpenAI and Microsoft in late 2022, are transformer-based language models that have caught the attention of researchers and lay-people alike but have yet to be applied to VHD research. The application of large language models to VHD and health care more broadly must be taken when using these limitations of these models as noted by the ChatGPT developers: “ChatGPT sometimes writes plausible-sounding but incorrect or nonsensical answers.”⁸⁰

The application of AI to VHD remains an emerging area of research and application development. The Food and Drug Administration has 3 specific pathways for the approval of AI/ML-based medical technologies and requires the use of “locked” algorithms—algorithms that remain static over time—for use as part of clinical care.⁸¹ However, beyond FDA approval, often several additional steps are necessary before broader adoption of the AI/ML based medical technology into clinical practice and incorporation into clinical guidelines. For example, the accurate identification of prevalent VHD with ECG or PCG represents an initial step. Prospective clinical trials evaluating the clinical effectiveness and cost-effectiveness of screening strategies are an important step for adoption into clinical guidelines. Furthermore, implementation studies examining barriers and facilitators to and costs of implementation in hospitals and clinics are also essential to broader adoption of these technologies.^{82, 83} Finally, several factors unrelated to the algorithms themselves (known as dataset shift) can lead to decreased model performance over time.⁸⁴ There remains a need for the development of robust governance systems to monitor model performance once deployed into clinical settings.

Strategies to better identify and reduce bias in ML models is an active area of investigation in the ML community. Several studies have shown that ML models can embed aspects of structural racism, gender, and implicit bias into their algorithms and potentially exacerbate health inequities.^85–88 Addressing these biases during the training, testing, and validation of AI/ML technologies for VHD will be critical to realize the potential benefit of these technologies across diverse populations and as a potential lever to advance equity.

In addition to concerns with biases in algorithms and deprecation in model performance when applied to new populations or over time, the “black box” nature of many algorithms remains a challenge to adoption. This issue is particularly salient in cardiovascular imaging as CNNs, which are black box algorithms, are widely used in studies. Although some clinicians may remain skeptical of AI-enabled technologies without the use of methods to identify the primary features contributing to a prediction, several experts have demonstrated the limitations of these methods.^{38, 39} Ultimately, similar to medications, the mechanism of action may be less important in the presence of high-quality evidence of benefit generated by randomized controlled trials. ³⁸

Guidelines are emerging to promote transparency in reporting and evaluating AI/ML clinical trials in a standardized fashion. The Consolidated Standards of Reporting Trials–Artificial Intelligence provides general recommendations for reporting of AI clinical trials in medicine that have been endorsed by several scientific journals.⁸⁹ The Proposed Recommendations for Cardiovascular Imaging-Related Machine Learning Evaluation (PRIME) checklist provides best practice strategies to promote standardization and reduce bias more specifically related to cardiovascular imaging.⁹⁰ Guidelines for the use of AI for echocardiography and other specific modalities are likely to develop in the near future in light of the growing impact of AI and myriad challenges, including clinical trial reporting standardization, bias reduction, and the understanding of medicolegal consequences of using AI/ML in practice.⁹¹ Adherence to reporting guidelines and sharing code and datasets when able would contribute to greater transparency, rigor, and reproducibility and acceleration research and application development in this area.^{90, 92–95}

In other areas of cardiovascular medicine, such as screening for atrial fibrillation or left ventricular dysfunction,^{37, 96–98} and in other specialties,⁹⁹ prospective cohort studies and clinical trials have helped form the foundation of the required evidence base for the scaling of these technologies. Similar studies in VHD will be essential to advancing the use of AI for improving outcomes for patients with VHD.

Summary

VHD is a morbid condition in which AI has the potential to improve screening, diagnosis, phenomapping, and risk prediction. Currently, most of the research and applications remain relatively early in the development and implementation lifecycle. Future research should prioritize adherence to reporting guidelines, sharing code and datasets when able, external validation in diverse populations, measuring and reducing bias in algorithms, prospective evaluation in clinical trials, and development of systems for ongoing monitoring of model performance.

Key Points.

• Valvular heart disease is a morbid condition with several unmet needs.

• Machine learning has the potential to improve the care of patients with valvular heart disease by facilitating screening, improving severity classification, identifying phenogroups, and improving risk prediction.

• Applications of machine learning in valvular heart disease are early in development and require additional testing.

Clinics Care Points.

• ML not only has the potential to improve the care of patients with VHD but also has pitfalls and limitations.

• ML approaches may in the future play an important role in the care of patients with VHD but AI-applications require rigorous testing, including prospective evaluation in clinical trials.