Artificial Intelligence in Congenital Heart Disease

The current era of big data offers a wealth of new opportunities for clinicians to leverage artificial intelligence to optimize care for pediatric and adult patients with a congenital heart disease. At present, there is a significant underutilization of artificial intelligence in the clinical setting for the diagnosis, prognosis, and management of congenital heart disease patients. This document is a call to action and will describe the current state of artificial intelligence in congenital heart disease, review challenges, discuss opportunities, and focus on the top priorities of artificial intelligence–based deployment in congenital heart disease.

However, there is a significant gap in the application of AI for diagnosis, prognosis, and management of CHD patients across their lifespan.The use of AI in pediatric and adult CHD has been limited by insufficient CHD-specific labeled data sets available for training of models, complex modeling needs in this patient population due to heterogenous clinical phenotypes and age-related pathophysiological changes, and siloed data in center-specific data warehouses. 8Additionally, at baseline, data for specific rare forms of CHD are limited, requiring multicentered collaboration to accrue sufficient data sets.Lastly, significant deficits in clinical training, knowledge, experience, and comfort with AI exist. 9spite these challenges, medical intelligence gained from the application of AI technologies and tools to data sets inclusive of the conglomerated CHD population could be instrumental in determining the optimal personalized management strategy for specific lesions.While some AI techniques currently used in adult cardiology may be transferable to adult CHD, 10 new techniques and collaboration are warranted to address the technical challenges specific to the complexity and rarity of CHD data sets.Therefore, strategic initiatives to promote AI-based research and clinical applications to best serve the unique needs of CHD patients are necessary.This document is a call to action and will describe the current state of AI in CHD, review challenges, discuss opportunities, and focus on the top priorities of AI-based deployment in CHD.

BASIC CONCEPTS OF AI
AI refers to any technique that enables computers to generate algorithms and find hidden insights to mimic human intelligence.Human intelligence is characterized by the ability to learn, reason, analyze, and make decisions.Machine learning (ML) is a subfield of AI that generates computer algorithms capable of improving task performance by learning or adapting from data. 8There are 3 ML strategies (Figure 1): 1) supervised; 2) unsupervised; and 3) reinforcement learning.Supervised learning uses labeled data sets to classify data or perform predictions. 8The goal of supervised learning is to learn a function from labelled data sets and produce desired outputs that best describes the relationship between the two. 6,8Unsupervised learning discovers the underlying structure or relationships among variables in an unlabeled data set without dependent variables. 6inforcement learning is determining the optimal behavior in an environment to earn the maximum reward and is the science behind decision-making.
Deep learning (DL) is a subset of ML that mimics the activity of the layers of neural networks in the neocortex.It has been used extensively in the field of medicine particularly for medical imaging using convolutional neural networks (CNNs), a specific type of deep neural network optimized for image analysis.More recently, CNNs have been applied to cardiovascular data sets for CHD. 1,11,12CNN models for image analysis are trained using raw imaging data sets and require substantial input data, computational power, and manual labor to label data.Transfer learning is an emerging approach that reduces the computational power needed and al- competing against each other to make more accurate predictions.This type of network is unsupervised and used in image generation. 13cent advances in AI include federated learning and swarm learning frame which can help with data privacy.Federated learning is an ML technique that trains an algorithm across multiple decentralized servers holding local data samples without sharing them. 14This type of learning allows the local devices attain the power to learning collaboratively from a shared model.After individual training of models on isolated data sets housed locally, the devices send their specific models to a centralized server where the models are averaged to obtain a single combined model.This process is repeated until a single highquality model is obtained.Swarm learning is a decentralized, privacy-preserving ML framework that does not rely on a central server. 15This type of framework uses the computing power at the distributed data sources to run the ML algorithms that train the model while maintaining data confidentiality. 15 such, these systems facilitate data sharing between medical centers.The next section describes the ML and DL models used in CHD.PRENATAL CHD SCREENING.7][18][19][20] Although fetal echocardiography in experienced hands has moderate sensitivity and high specificity reported in the recent meta-analyses, the accuracy of CHD detection is reported as low as 28% in general obstetric practice. 21,22e acquisition of standard cardiac imaging planes is critical in the prenatal diagnosis of CHD, and using AI to automatically retrieve these standard imaging planes from a stream of ultrasound imaging data has the potential to improve CHD detection.The automatically retrieved images may be of higher quality than the manually obtained images.Baumgartner et al 23 used labeled mid-trimester ultrasound images from 2,694 volunteers and a CNN algorithm to achieve real-time classification of standard-screening fetal cardiac imaging planes (Table 1).Dong et al 25  POSTNATAL CHD SCREENING.Initial screening in infants for CHD consists of a combination of cardiac auscultation, pulse oximetry, chest radiography, and electrocardiography.9][50] Gharehbaghi et al 27 described using combined support vector machine and hidden Markov models-supervised learning models used for classification-to identify innocent heart murmurs from a bicuspid aortic valve with an accuracy of 86.4%, better than pediatric cardiologists using conventional auscultation (Table 1).This software has not been accepted for clinical use because it was limited by small numbers and lack of widespread clinical deployment. 27Furthermore, it was also unclear how the data were split between training and testing data sets. 27A subsequent study by Gharehbaghi et al 28 1).The heart sounds were preprocessed and segmented, then followed by feature extraction.The features were fed into a boosted decision tree classifier to estimate the probability of patent ductus arteriosus or CHD from normal heart sounds. 30Finally, the patients were prioritized into the decision of getting echocardiograms to confirm the diagnosis. 30This study was the first to identify patent ductus arteriosus using a designed ML-based method and contrast it with an experienced neonatologist's auscultation skills as well as the gold standard of echocardiogram.The model area under the curve was 0.77 for the detection of patent ductus arteriosus. 30The authors suggested integrating pulse oximetry to the ML algorithms could improve their framework of a more comprehensive assessment of the performance of AIaugmented decision-making as a clinical decision support tool; however, there has yet to be wide uptake in clinical practice. 30Smart stethoscopes have been developed to help with clinical auscultation for detecting CHD heart sounds in remote areas where resources and pediatric cardiology expertise are limited. 51Quickly identifying abnormal heart sounds may help triage these patients with appropriate referral for CHD management. 48DL has been applied to chest radiographs of CHD patients to aid in the prediction of pulmonary-to-systemic flow ratio (Table 1). 29DL has also been applied to electrocardiogram (EKG) readings to detect atrial septal defects. 31This DL algorithm that comprised CNN and long short-term memory models may be applicable to other CHD lesions to further aid EKG-based AI diagnostics for the pediatric population.
Wearable technology has gained rapid acceptances into the pediatric community for the analysis of heart rate, blood pressure, oxygen saturation, and heart rhythm.These technologies have the potential of lifesaving monitoring in the outpatient setting or remote monitoring for children with CHD and arrhythmias.
ML interpretation and prediction using EKG input data are being used for the detection of CHD (Table 1). 31The described algorithms for pediatric arrhythmia detection include those for smartwatches 52 and zio patch 53,54 devices.More recently, an ML algorithm built from pulse oximetry features has been created to improve critical CHD detection rates (Table 1). 32Future embedding of AI algorithms into the wearable technologies will help to develop connected intelligence, early warning systems, which can be used for prompt risk stratification, targeted early intervention, and personalized prescription.
These devices can be used as predictive devices rather than diagnostics.

Small data sets
Continued on the next page Jone et al Artificial Intelligence in Congenital Heart Disease   1), 13,37 echocardiography, 63 and cardiac CT. 64 Tandon et al 37 showed that a CNN algorithm for CMR, developed for structurally normal hearts, was able to be adapted to use in a repaired tetralogy of Fallot (ToF) heart with a relatively small number of training data sets (Table 1).They proposed that similar work can be extended to other forms of CHD.
Automating an otherwise manual step could Preprocedural planning in catheterization and CHD surgeries requires cardiac imaging integration.GANs (a type of neural network that learns to generate new data from training data sets) have been used successfully to predict the optimal size, shape, and positioning of the transannular patch to optimize outcomes from cardiac CT images of ToF patients. 69 a pilot study, cycle adversarial networks were able to align preprocedural CTs with intraprocedural transesophageal echocardiographic images to improve surgical navigation for patients with CHD. 40iz-Fernandez et al 39 optimized AI-based algorithms to improve risk estimation for CHD surgery (Table 1).
Deploying AI to automatically segment the pulmonary veins and the left atrium prior to total anomalous pulmonary venous return repair can be crucial for presurgical planning. 64Integrating AI into virtual reality in the repair of atrioventricular septal defect using 3D echocardiographic imaging may help with surgical repairs. 70 The feasibility of AI-derived automated risk stratification has been demonstrated.Based on raw medical record data from over 10,000 adult CHD patients, DL algorithms using natural language processing can determine the underlying complexity of disease and predict the need for closer medical attention as well as the risk of mortality in this population. 10Furthermore, a study from the German National Register for CHD has shown that AI-based direct risk stratification can be achieved using raw cardiac CMR from patients with ToF.This study is also one of few to assess the external validity of the algorithms by geographically separating training and testing data sets. 71 Integrating multiple sources of medical data for a multimodal approach has also enabled prediction of mortality in patients in the intensive care unit. 80ese multimodal models take advantage of complex inputs such as electronic health record with a wide variety of data such as medical diagnoses, vital signs, prescriptions, and laboratory results to make predictions similar to human clinicians making decisions based on diverse information in clinical practice.Therefore, this multimodal AI approach is a unique opportunity to expand to CHD data sources, which are known to be quite varied.shown the feasibility of CHD screening using serum metabolite panels. 81Furthermore, DNA methylation can be used to predict aortic coarctation in neonates (Table 1). 46A clustering analysis from unsupervised learning might uncover new subtypes of patients that could benefit from a similar treatment or management.Cainelli et al 82 applied clustering analysis to children with CHD who had underwent a cardiac surgery to discover 2 distinct profiles: those with a high burden of psychopathology and those with similarities to patients with attention deficit hyperactivity disorders.
Digital twin technology simulates a vision of a comprehensive virtual tool that integrates dynamic clinical data of a patient over time using a mechanistic and statistical model. 83The technology serves as a real-time counterpart for a patient and uses mobile health-monitoring data, "omics", clinical reports, clinical and experimental recordings, and 5][86] Real-world data can be continuously fed into models to arrive at better prediction outcomes rather than relying on registries or randomized control trials.AI will assist in precision Creating a CHD consortia, standardizing data with multicenter data collection, and using federated learning are proposed solutions to overcome heterogeneous data sets. 5,8In many circumstances, investigators will have to rely on minimal common data sets, focused on data that are homogeneous across most or all centers, to base the core of the analysis on these data, while using data that are more subject to variability only for secondary analyses.Moving forward, it will be important to increase the standardization of study protocols, data collection, and the homogenization of the definition of clinical findings and symptoms, ideally collected with standardized forms.For example, the use of digital imaging and communications in medicine has facilitated the standardization of imaging in medicine, as standard formats for EKG raw signals have facilitated multicenter AI research using EKG. 93,94Similarly, the investigators will need to identify ways to make electronic health record interoperable, something that will facilitate the collection of clinical data from multiple centers.
DATA IMBALANCES AND BIAS.Data sets in CHD, particularly medical imaging information, are subject to imbalances in representation that can be perpetuated by AI development.For example, CMR data sets contain primarily abnormal subjects with very few normal cases, and the opposite is true for echocardiography.Small and single-center data sets can lead to skewed and overfitted training data sets to specific populations, limiting applicability to real-life scenarios such as patients of different races or socioeconomic classes. 95,96nderrepresentation in available data sets may limit access to AI-driven solutions, compounding health care inequity in the CHD population. 97,98 address the issue of data imbalance, the CHD community will need to create prospective data sets of normal children in research setting to have a point of comparison for when the algorithm requires normal controls for echocardiograms, EKGs, CTs, or other cardiac testing.ML solutions will also need to be tested rigorously in different settings and populations so equity and the potential for bias can be continuously monitored after implementation.To address the "explainability" issue, computer scientists are currently working on methods to identify the key explanatory variables, in order to decode the "black box."Those efforts will help to ease the concerns of knowing exactly how the computer identifies specific conditions or predicts outcomes.
However, some degree of uncertainty in how the AIbased solution works will have to be accepted.This These deidentified data sets can be housed in the cloud so that there is a common repository of these CHD lesions for multiple centers to access to produce meaningful AI solutions.As such, expansion of training data sets with rare CHD cases can be achieved.This would allow for dissemination of expertise in diagnosis and management of CHD lesions. 99e feasibility of progressive GANs has been demonstrated in CMR. 90 lows for faster training of the model.It uses pretrained CNN weights to extract features to apply to the new CNN model to reduce the amount of training data needed to build the model.Another form of DL is recurrent neural networks (RNNs) which use the outputs of some layers of neural network as feedback to use as inputs to the previous layer.This allows for sequential data analysis.A popular framework for learning sequential data is called the long short-term memory network.This is a type of RNN that is capable of learning long-term relationships.Transformers, another type of DL model that learns context and tracks relationships in sequential data, are primarily used in the field of natural language processing and computer vision.RNNs based on long short-term memory or grated recurrent units are capable of learning information dependencies in long input sequences, but they are not parallelizable because the hidden states are computed sequentially.Instead of relying on recurrent structures, transformers process the entire HIGHLIGHTS Significant unmet opportunities exist for artificial intelligence to advance congenital heart disease research.Leveraging artificial intelligence in congenital heart disease could reduce repetitive tasks and augment clinical decision-making.Future research in artificial intelligence should include longitudinal congenital heart disease data to map disease pathophysiology and prognosis across the lifespan.A B B R E V I A T I O N S A N D A C R O N Y M S AI = artificial intelligence CHD = congenital heart disease CMR = cardiac magnetic resonance imaging CNN = convolutional neural network CT = computed tomography DL = deep learning EKG = electrocardiogram ML = machine learning RNN = recurrent neural network input using a self-attention mechanism, which is easy to be computed in parallel.One potential limitation of transformers is that the computation is very memory-intensive.Transformers in AI are the most recent advancement with implications for CHD clinical work.A generative adversarial network (GAN) is a DL model involving 2 neural networks CURRENT AI-BASED PEDIATRIC AND ADULT CHD APPLICATIONS AND OPPORTUNITIES Over the last decade, there has been an exponential rise in the number of publications centered on AI in health care, highlighting the potential of this technology.Applications of AI for CHD are robust, ranging from prenatal screening to risk stratification in an aging adult CHD population.In the following section, we review the progress in this field and highlight opportunities to advance unmet needs in areas of prenatal CHD screening, postnatal CHD screening, cardiac imaging processing and interpretation, preprocedural planning, outcome prediction, and precision medicine (Central Illustration).
used a CNN in 2,032 fetal 4-chamber views and 5,000 views of other fetal structures to evaluate for automatic detection of the 4-chamber view and automatic assessment of image quality.Chen et al 24 used a composite RNN model in an ultrasound scan of 1,231 ultrasound videos of fetuses to automatically detect standard planes including 4-chamber cardiac views.A landmark article by Arnaout et al 12 using CNN described the ability to detect complex CHD in utero from normal fetuses.These algorithms can help clinicians and operators who are less experienced in evaluating fetal echocardiograms to detect abnormalities to improve the detection rates of CHD in the community.This could also reduce work time required to obtain normal standard views and allow the sonographer to use the retained time to focus on the evaluation of abnormal cardiac pathology. 47Further refinements in AI algorithms or development of fetal CHD-specific learning algorithms could help achieve more granular detections of unique CHD lesions.This has the potential to risk stratify certain fetal populations.Examples of this include improved detection of patients at risk of ductal dependent physiology (ie, coarctation of the aorta) or at risk of intrauterine fetal demise.
improve the precision of measurements and efficiency of interpretation.With the implementation of AI, one can reduce the amount of time required in acquiring images, processing images, and reducing variability in interpretation of the images.Computer vision can be leveraged to streamline imaging evaluation and interpretation.By using AI algorithms such as deep neural networks, clinicians can analyze largevolume, nonnumerical data structures such as image processing and apply them to multi-imaging evaluation.This has already been successful in cases of noncongenital cardiac diseases, but there has been limited application in the evaluation of congenital cardiac pathology. 65Through the development of CHD-specific learning algorithms, AI could shorten the image-acquisition time, improve image processing, derive interpretation, and facilitate a prompt and precise diagnosis.Future opportunities for AIenabled echocardiograms from image acquisitions to image interpretation can be developed.In particular, AI algorithms that work with limited labeled data using novel self-supervised and semisupervised approaches will be helpful in CHD.Implementation of AI in CMR evaluation could significantly benefit the pediatric CHD community.Further development of DL algorithms for CMR reconstruction has the potential to reduce CMR scan time and minimize the effects of motion artifact on imaging quality. 13,65-68This could have the benefit of minimizing the need of anesthesia for CMR evaluation of young or uncooperative pediatric patients.Furthermore, this could be an opportunity to accelerate fetal CMR research and development.Overall, DL algorithms have the potential to reduce postprocessing time for both the technician obtaining the images and the physician interpreting the study.PREPROCEDURAL AND PRESURGICAL PLANNING.
PRECISION MEDICINE.AI in precision medicine involves taking multiple large data points in CHD patients to arrive at a more-accurate diagnosis and specific CHD phenotypes.Big data analytics will help drive the individualized therapy and interventions required for CHD patients.One considerable advantage of AI is its ability to serve as a tool to aggregate and synthesize the many layers of medical data to offer personalized analytics.This includes clinical information, environmental factors, imaging data, and social determinants.Genomic medicine derived from AI will allow for better characterization of the underlying pathophysiology of CHD.Studies have medicine by taking multiple large data points for CHD patients to arrive at more accurate and specific diagnoses of CHD phenotypes.Genomic medicine derived from AI will allow for better characterization of the underlying pathophysiology of CHD.Big data analytics will help drive the individualized therapy and interventions required for CHD patients.CHALLENGES IN IMPLEMENTING AI IN CHD AND PROPOSED SOLUTIONS Challenges exist in implementing AI in CHD.The first essential step to bringing the potential of AI to reality is to identify significant barriers to integration.Substantial issues related to AI development and integration include the lack of adequate education about AI for clinicians, low volume of data, heterogeneity of data, data imbalance, "explainability" of AI models with interpretability of data, need for collaboration between clinicians and data scientists, and legal barriers.In the following sections, we identify the challenges and propose solutions.LACK OF AI EDUCATION FOR CLINICIANS.The lack of adequate AI education for clinicians poses significant challenges in the clinicians' understanding of AI and subsequent adoption within clinical practice.As with any new modality in medicine, training and understanding of the language used in that modality must be met in order for clinicians to implement this new modality in clinical practice.Without adequate education in AI, it is difficult for clinicians to work with data scientists to create a meaningful clinical project that would be useful for CHD.One proposed solution is introducing AI education in the medical education curriculum and integrating it into the categorical pediatric cardiology fellowship so that the next generation of clinicians will have adequate understanding of AI.LOW VOLUME OF DATA.Unlike conventional ML models, DL typically requires large data sets for model training.The complexity of CHD and heterogeneity within lesions pose a challenge to collect sufficient data sets that are representative of the breadth of the disease for reliable AI model development.Overfitting algorithms to single-center data sets due to insufficient data and lack of external validation remains a large barrier to gaining physician buy-in, acceleration of research initiatives, and wide implementation of developed algorithms into the clinical environment.Networks for data collection in CHD have been developed, including, but not limited to, ACTION (Advanced Cardiac Therapies ImprovingOutcomes Network),87 FON (Fontan Outcomes Network),88 and PROTEA (PartneRships in cOngen-iTal hEart disease).89Nonetheless, even these highly curated retrospective data sets are often unable to handle the noisy artifact-laden data generated during patient care.When data are clean and available, such as robust echocardiography data sets, it often requires manual labeling or input for training algorithms, which entails a significant time burden.To overcome a small sample size in CHD, Diller et al 90 used a strategy by generating 100,000 synthetic images based on CMR data from 303 patients with ToF deemed anatomically plausible by human observers and achieved similar results in comparison to the original patient data.Synthetic data generation, while potentially useful, may lend to modeling bias.Another proposed solution is using GANs or other methods such as sampling for augmenting data from small data sets or some data-insufficient applications.Lastly, some researchers have used transfer learning to overcome data size limits by leveraging model parameters trained on larger data sets.This may be a viable solution in addition to multicenter collaboration.HETEROGENEITY IN DATA AND CURATION.Patients with CHD have a wealth of data sets from wearable devices, intensive care unit stays, and imaging data from multimodality studies.Systems for recording an accurate alignment of events in time might differ by institutional standards.While the abundance of data provides a favorable foundation for algorithmic development, these heterogenous data sets often reside in disparate repositories and formats, creating barriers to access and multicenter collaboration.For example, in CMR, there are different commercialized CMR machine vendors, protocols between hospitals, and varied storage systems.Moreover, a mismatch in data due to a change of environment or disease stage between training and operational data can result in erroneous predictions.This issue has been compounded by the rise of wearable technology with different manufacturers and proprietary data formats.Even in the presence of homogeneity in vendor data sets, hospital infrastructures are poorly equipped for large-scale algorithm development.As a result, multicenter uniform standards and vendor-agnostic model development are necessary to mitigate the heterogeneity of the scanned data.91,92 EVIDENCE AND "EXPLAINABILITY"-TRUST.The high stakes of managing CHD warrant evidence and "explainability" of research, particularly to overcome the reluctance among clinicians unfamiliar with AI to adopt the technology in clinical practice.The majority of clinicians are experienced with traditional and transparent health care research models.In contrast, in AI model development, algorithms may be programmed to arrive at the output without clear instructions; the infamous "black box" problem.The inability of the AI system to explain how it arrives at the prediction is a serious technical challenge that prevents trust from clinician users.Furthermore, reports of publicly available AI tools potentially causing harm to patients display the potential downfalls of AI-based solutions used without appropriate validation.As such, clinicians and the public will likely mandate a degree of "explainability" before AI integration.It is also unlikely that clinicians will trust the AI model if it does not give correct predictions.
is akin to the number of medications used in clinical practice for decades which have no clear or known mechanism of action but are used routinely.To this end, as long as the AI-based solutions prove to be safe and effective to detect conditions or predict outcomes, clinicians, patients, and the public in general will become more tolerant to the relative uncertainty to understand how the machine works.NEED FOR COLLABORATION BETWEEN CLINICIANS AND DATA SCIENTISTS.A major challenge in integrating AI with medicine has been the disconnect between clinical investigators and computer scientists in terms of what is important for patients, the definition of problems, and ways to solve them using AI.To address this, engineers and computer scientists will need to become more familiar with clinical practice and to see firsthand the potential AI-based solutions that would benefit clinicians and patients.At the same time, clinicians and scientists without prior experience in computer sciences will need to embrace new information and skills to better understand the way ML is developed.Over time, people from both sides will also need to appreciate that the terminologies that are often named differently actually refer to the same concept.For example, what clinical investigators call variables, computer scientists call features, and what clinical investigators call outcomes, engineers call labels.Building bridges in education between both fields is imperative to accelerate the development of clinically useful tools.LIABILITY AND LEGAL CONCERNS.Members of the hospital legal system are often unfamiliar with how to address accountability and liability should health care professionals utilize AI in practice.Incorrect predictions made by AI algorithms can result in severe, lifelong consequences for patients, requiring a high degree of caution, oversight, and quality control.Ambiguity in terms of intellectual property and who ultimately owns the data (ie, patients, the hospital, or developers) may also complicate the milieu.As with any innovation, legal corollaries regarding the use of patient-generated data for partnerships with industry to develop AI-based systems must be determined.Currently the American Medical Association and the Food and Drug Administration are independently working on defining major ethical and legal dilemmas brought by AI in medicine.It is expected that national scientific societies, federal agencies, and other organizations will come up with clearer guidelines addressing those potential legal and ethical dilemmas in the near future.CALL TO ACTION IN CHD A call to action to broaden the expansion of AI in CHD requires multicenter collaboration, curation and creation of data sets, building institutional AI infrastructure, and implementing AI best practice and AI education and training.MULTICENTER COLLABORATION.Multicenter collaboration is necessary to accrue large data sets to train AI algorithms.Multicenter registries have been started.

Artificial
Key areas to focus in future AI research and deployment in CHD include: 1) targeting rare diseases (coronary artery anomalies); 2) acquired diseases that disproportionately affect patients with limited access to pediatric cardiologists for disease diagnosis (rheumatic heart disease); 3) congenital heart anomalies with high morbidity and mortality (failing Fontan or ToF patients at risk of sudden cardiac death); and 4) precision medicine for decision-making in difficult diseases (borderline left heart).Randomized control trials are difficult to perform in CHD because of the disease complexity, disease rarity, and clinical heterogeneity of different lesions.Together, these decrease the precision of the treatment options for patients as most CHD treatment recommendations are based on expert consensus.Collaborative AI research focused on aggregating data and sharing insights into rare diseases that could help clinicians with better decision-making.Some areas of urgent need for AI research have been explored and contain ample opportunities for future collaboration and extension.For example, Meza et al 44 used unsupervised ML to identify patterns in echocardiographic data that could be clinically relevant to diagnosis and prognosis of patients with borderline left ventricle.In this study, parameters like mitral valve characteristics and pulmonary vein anomalies that are often used in clinical practice to help guide management in the borderline left ventricular condition were not found to be significant in distinguishing patients from 3 different groups (multilevel left ventricular hypoplasia, hypoplastic left heart syndrome, and critical aortic stenosis).Diller et al 71 developed an automatic DL imaging algorithm that predicted death/aborted cardiac arrest and documented ventricular tachycardia in ToF patients.Other applications such as the prediction of the feasibility of and risk associated with surgical or catheter-mediated interventions 42,43,69,72,101 and individualized prediction of drug effects or interventions in complex hemodynamic settings show promise for future CHD extension. 86With the advent of increased computing power, clinicians can leverage AI for precision medicine with better clinical decision-making, and patients can receive real-time information about their personal health metrics.With increased cognitive computing with natural language processing, reinforcement learning, and DL, AI will have better future prediction models and drug therapeutics in patients with CHD.Digital twin will provide cardiologists the best therapy without relying on published reports or registries in the future.CONCLUSIONS The unique strength of AI models is the uncanny ability to learn from data with increased exposure.Leveraging AI to accelerate and strengthen CHD research and clinical applications is now possible largely due to the escalating volume and complexity of data available and advent of increased computing power.Clinicians and patients could soon benefit from clinical decision-support tools that assist with personalizing patients' diagnosis, prognosis, and treatments and provide real-time information on J A C C : A D V A N C E S , V O L . 1 , N O . 5 Intelligence in Congenital Heart Disease personal health metrics.Although there are challenges in the implementation of AI in CHD, opportunities exist in many areas of CHD for clinicians to explore.With the arrival of newer AI-powered algorithms capable of handling big data, such as DL using CNN and RNN, federated learning, and digital twin, a vast amount of research opportunities exist to collaborate and study CHD across a lifespan to build future prediction models and develop drug therapeutics in patients with CHD.

TABLE 1
Application of Artificial Intelligence in Congenital Heart Disease Continued on the next page CARDIOVASCULAR IMAGE PROCESSING AND INTERPRETATION.Cardiac imaging such as echocardiography, cardiac magnetic resonance imaging (CMR), and computed tomography (CT) serve as the core of diagnosis and disease surveillance but require significant expertise and time for acquisition and interpretation.DL has been applied to improve each stage of multimodality imaging acquisition and interpretation in adult cardiology (preprocessing, quality optimization, view classification, segmentation, and diagnosis) with less advancement in pediatrics. 55-61This is in large part due to the breadth and subtleties of disease and fewer available training data sets for complex CHD patients, which limits the performance of DL models, increases the chances of overfitting data, and limits the opportunities to externally validate models.Because of this, most published studies focus on specific diseases, making

TABLE 1 Continued
35,tinued on the next page it necessary to implement algorithms not generalizable to the spectrum of CHD into clinical practice.For example, Diller et al62built a CNN algorithm capable of discriminating echocardiograms in adult CHD patients with transposition of the great arteries after the atrial switch, patients with congenital corrected transposition of the great arteries, and healthy controls (Table1).Although this study had 98% accuracy in identifying CHD, the CNN algorithm has not been externally validated for clinical deployment.Echocardiogram clips have been used to train DL models for the automated diagnosis of atrial septal defects, ventricular septal defects, and coarctation of the aorta.35,38Furthermore,DL has been used for the

TABLE 2
Barriers to Artificial Intelligence Implementation in Congenital Heart Disease and Proposed Strategies to Overcome Them 1. Insufficient data access, storage, and sharing strategies for CHD patient data.Data limitations are due to lack of accurately labelled data.Methods such as transfer learning, self-supervised learning, and predictive learning to increase these data may help overcome these barriers to increase opportunities for external validation 2. Lack of AI in medicine awareness from stake holders in health care (ie, clinicians, patients, and hospital administrators).Clinicians need more education about data and AI, and patients need more education to understand the need for and benefits of collaboration on real-world data and not just registries and randomized control trials.Developing institutional educational series and profession society webinars (American College of Cardiology Innovation and Adult Congenital and Pediatric Cardiology sections) may help address these challenges.3.Absence of forums to facilitate communication between clinicians and data scientists.Providing computer and data scientists with more knowledge regarding the proposed deficits in health care to target the development of meaningful AI solutions.Increasing clinician-to-data scientist synergy for mutual understanding of the dual perspectives of both domains.4. Difficulty harnessing collaboration.Recruitment of multidisciplinary team members, particularly AI champions, to drive AI implementation.5. Current CHD research is unidimensional.Leveraging multimodal AI for cardiology to incorporate the full spectrum of data: genomics, imaging, demographic, ICU, wearable, and so on to accelerate precision medicine.6. Concern that AI methods are not transparent enough for the medical community.Utilizing explainable AI to minimize the "black box" perception of AI and requiring studies provide documentation that they completed the recommended Minimum Information About Clinical Artificial Intelligence Modeling Checklist.7. Critique that AI projects are not created in the context of clinical applicability.Utilizing design thinking to select proper AI methodology relevant to the clinical context.8. Poor acceptance of AI in the research community and concern that AI requires too much time to establish sufficiently large data sets.Using innovative AI methods to leverage the power of small data sets.Executing more realistic projects that are easier to accomplish, with demonstrable value and return on investment (ROI) may help get "buy-in" from the administrative and clinical leadership.
AI ¼ artificial intelligence; CHD ¼ congenital heart disease; ICU ¼ intensive care unit.CHD AREAS FOR IMPLEMENTING AI: TOP PRIORITIES.