Role of artificial intelligence in congenital heart disease

Abstract

This mini-review explores the transformative potential of artificial intelligence (AI) in improving the diagnosis, management, and long-term care of congenital heart diseases (CHDs). AI offers significant advancements across the spectrum of CHD care, from prenatal screening to postnatal management and long-term monitoring. Using AI algorithms, enhanced fetal echocardiography, and genetic tests improves prenatal diagnosis and risk stratification. Postnatally, AI revolutionizes diagnostic imaging analysis, providing more accurate and efficient identification of CHD subtypes and severity. Compared with traditional methods, advanced signal processing techniques enable a more precise assessment of hemodynamic parameters. AI-driven decision support systems tailor treatment strategies, thereby optimizing therapeutic interventions and predicting patient outcomes with greater accuracy. This personalized approach leads to better clinical outcomes and reduced morbidity. Furthermore, AI-enabled remote monitoring and wearable devices facilitate ongoing surveillance, thereby enabling early detection of complications and provision of prompt interventions. This continuous monitoring is crucial in the immediate postoperative period and throughout the patient’s life. Despite the immense potential of AI, challenges remain. These include the need for standardized datasets, the development of transparent and understandable AI algorithms, ethical considerations, and seamless integration into existing clinical workflows. Overcoming these obstacles through collaborative data sharing and responsible implementation will unlock the full potential of AI to improve the lives of patients with CHD, ultimately leading to better patient outcomes and improved quality of life.

Key Words: Heart diseases; Congenital abnormalities; Artificial intelligence; Echocardiography; Diagnosis

Core Tip: Artificial intelligence (AI) offers transformative potential for congenital heart disease (CHD) care, affecting diagnosis, management, and long-term monitoring. This study explores the multifaceted applications of AI across the journey of patients with CHD, highlighting key advancements and critical challenges. Prenatally, AI-enhanced fetal echocardiography and genetic testing promise earlier and more accurate diagnosis, allowing for timely intervention. Postnatally, AI-driven image analysis accelerates diagnosis, and advanced signal processing improves hemodynamic assessment. AI-driven decision support systems tailor treatment strategies based on individual patient characteristics. Long-term care benefits from AI-enabled remote monitoring and wearable technologies, facilitating proactive management and early detection of complications. However, realizing the full potential of AI in CHD requires addressing significant limitations. The development of robust, standardized datasets is crucial for training reliable AI models. Furthermore, ensuring the transparency and explainability of AI algorithms is essential for building trust and accountability. Ethical considerations, including data privacy, bias mitigation, and equitable access, must be carefully addressed. Finally, the seamless integration of AI tools into existing clinical workflows is vital for practical implementation and widespread acceptance. Addressing these challenges will pave the way for AI to revolutionize CHD care, achieve better outcomes, and improve the lives of patients and their families.

INTRODUCTION

Congenital heart diseases (CHDs) account for nearly one-third of all major congenital anomalies, representing a significant global health challenge[1]. These defects can be life-threatening at birth or later, with symptoms such as failure to thrive, heart failure, or cyanosis. CHDs exhibit significant heterogeneity in their genetics, pathophysiology, clinical presentation, course over time, and treatment, necessitating specialized care, one or more corrective or palliative surgeries or interventions, and lifelong medical follow-up, imposing a substantial care burden on families and healthcare systems[2,3]. Despite advancements in medical care that have dramatically improved survival rates, significant challenges remain in early diagnosis, treatment planning, and long-term management of CHDs.

Rapid advances in artificial intelligence (AI) have recently caused disruptive changes to most domains of human activity. AI refers to the ability of computers to imitate or simulate intelligent human behavior or accomplish tasks that typically require human intelligence, including learning, comprehension, reasoning, problem-solving, and decision-making, even creativity and autonomy. Appropriate utilization of AI may transform CHD care and management; however, these possibilities are not without risks, challenges, and dangers. We will attempt to explore these advances, ideas, and possibilities, as well as the dangers and precautions, in this review to orient and help practitioners navigate the rapidly evolving face of medical care.

The steps in CHD care can be loosely divided into prenatal diagnosis, postnatal diagnosis, medical management, intervention, surgery and perioperative care, and long-term care[3].

PRENATAL DIAGNOSIS

Prenatal diagnosis of CHDs improves survival, allowing informed decision-making regarding the continuation of the pregnancy, postnatal management, and in utero fetal interventions[4,5]. Fetal echocardiography is the mainstay of prenatal diagnosis. Although the optimal accuracy of fetal echo with experienced practitioners is > 90%, real-world screening yields are variable and, on average, much lower at approximately 30%[6]. AI-based solutions have been explored to help in image processing, biometric measurements, and automatic disease diagnosis[7]. Fetal ultrasonography (US) is complicated because of the complex anatomy, movements, artifacts, and quick heartbeat of the fetus; however, deep learning techniques such as convolutional neural networks (CNNs) can identify important features and categorize scan images with up to 95% accuracy[8,9]. Similarly, algorithms have been demonstrated to measure quantitative parameters such as volumes, ejection fractions, atrioventricular plane displacements, and fetal cardiac systolic time intervals[10,11]. Thus, derived biometric measurements can predict the coarctation of the aorta in utero with an area under the receiver operating characteristic curve (AUROC) of 0.96[12]. Finally, AI models have also been trained to directly identify pathology from fetal US images, giving an AUROC of up to 0.99[13,14]. However, these measures of accuracy stem from validation in a testing dataset and not from real-world use. Thus, this question remains: How does AI-enabled fetal echocardiography compare with the current standard of care? Interestingly, Day et al[15] compared the performance of optimally trained AI models with the current (2024) national screening program for hypoplastic left heart syndrome (HLHS). The Fetal Anomaly Screening Programme of the United Kingdom achieved 94.3% sensitivity and 99.985% specificity from a fetal echo in the second trimester. The performance of an AI model trained on the data depended on calibration. If calibrated for sensitivity, it detected 14 additional HLHS cases at the expense of 45134 screen-positive results; if calibrated for specificity, it would detect two fewer cases with 118 fewer false positives[15]. This evidence indicates that human performance is still superior for qualitative imaging screening.

Genetic and chromosomal testing is the other main method of diagnosing prenatal disease. This process has also not been untouched by AI. AI can achieve basic tasks such as image segmentation for karyotyping and more complex ones[16,17]. Novel approaches based on machine learning on fetal fraction or cell-free DNA from noninvasively obtained maternal samples are also being investigated and may represent a promising area of study for prenatal diagnosis of CHDs[18-20].

POSTNATAL DIAGNOSIS

Postnatal diagnosis of CHDs is based on history, clinical examination, including auscultation, pulse oximetry, and investigations, such as electrocardiogram (ECG) and echocardiography.

Echocardiography serves as the cornerstone of the definitive diagnosis of CHDs. Similar to other imaging modalities, automation and machine learning are increasingly being incorporated into echocardiography. The extent of implementation of AI in echocardiography has already been reviewed. Accordingly, we have proposed four steps for AI interpretation and handling of echocardiographic images: Acquisition, chamber segmentation, chamber size and function determination, and abnormality detection[21].

Each step has undergone significant work. A deep learning-based tool has been demonstrated to guide novice echocardiographers to obtain standard views of diagnostic quality in > 90% of cases[22]. Such approaches can help in wider echocardiographic screening and surveillance for CHDs.

Availability of large labeled and unlabeled datasets has allowed the training of CNNs to identify and segment cardiac chambers from echocardiographic images with high accuracy[23,24]. Adaptive analytic algorithm software for segmentation and automatic ejection fraction determination from 3D echocardiographic images is valuable and has already been deployed in echo workstations[25]. They show high accuracy, esp. in right ventricle (RV) quantification, as they’re free from the geometric assumptions of conventional echocardiography. Clinical studies in adults show high accuracy, reproducibility, and time efficiency, and they’re already recommended as standards of care in guidelines. However, while this works well in children above 6 years, usage in CHD cohorts with younger children and variable cardiac morphology would need further training and validation[26].

Even more intriguing are cases where AI is improving the capability of imaging modality, as is seen in magnetic resonance imaging (MRI). Deep learning-based reconstruction methods in MRI have attracted interest due to their potential to reduce scan times massively[27]. Cardiac MRI offers useful information; however, it can be difficult for children with CHD because of the long scan time and the potential need for sedation. Thus, AI-based fast MRI would decrease the risk and improve the adoption of cardiac MRI in CHD[28].

However, echocardiography requires significant labor and technology, and screening and surveillance of CHD are more important. Thus, the diagnosis of CHD using simpler methods such as auscultation or ECG offers more promise for public health.

Decision tree classifiers or deep learning-based approaches have been used to diagnose CHDs, such as ventricular septal defects or patent ductus arteriosus from phonocardiograms[29,30]. A virtual test using AI algorithms to detect heart murmurs, based on recordings from the Johns Hopkins Cardiac Auscultatory Recording Database, matched the best standard for heart rate detection, with 93% sensitivity (CI, 90%-95%), 81% specificity (CI, 75%-85%), and 88% accuracy (CI, 85%-91%) for murmurs, indicating strong potential for the future[31].

The advantage of AI is deep feature extraction, allowing inferences not offered by conventional clinical methods. Mori et al[32] demonstrated the prediction of atrial septal defects using a deep learning method with a CNN and long short-term memory, demonstrating sensitivity, specificity, and accuracy of 0.76, 0.96, and 0.89, respectively. Interestingly, this was superior to board-certified pediatric cardiologists based on the same data. Similarly, deep neural networks showed better discrimination than QTc in diagnosing long QT syndrome[33]. Even more interestingly, Mayourian et al[34] trained a CNN model on paired ECG and cardiac MRI data from a heterogeneous cohort of children with various CHD. Surprisingly, the model could predict left ventricular (LV) and RV dilatation and outcomes such as LV and RV dysfunction with good discrimination (AUROC, 0.89 and 0.82, respectively)[34]. Deep learning from chest X-ray images for pulmonary-to-systemic flow ratio has displayed a better correlation with catheterization-derived values than expert inference, reinforcing the wealth of incremental information available through AI[35].

MANAGEMENT AND DECISION-MAKING

Management and decision-making in CHD rely on guidelines, experience, and understanding of the disease; however, the rarity and variability of diseases make the use of evidence-based medicine challenging[36-39]. An intriguing concept advocated in such scenarios is “medicine-based evidence”[40]. It involves customizing medical decision-making for a patient with a specific clinical condition by integrating all available data regarding past medical interventions in comparable real-world patients. Potentially, this framework allows the synthesis of multiple layers of data—from history, clinical data, laboratory results, and imaging to the patient’s environment, microbiomics, proteomics, metabolomics, genomics, epigenomics, and transcriptomics. At present, this paradigm has become feasible by the wealth of clinically available data, as well as AI tools such as deep learning, unsupervised learning, and reinforcement learning[41]. These tools can facilitate the development of a “digital twin” of a specific patient, which is a prediction model that incorporates their unique anatomy, pathophysiology, and genetics, along with mechanistic and statistical modeling. This model allows for predictions about how the patient would respond to proposed therapeutics[42]. This paradigm appears to be an attractive option for precision medicine in the anatomical, physiological, and phenotypic diversity of CHDs[43]. Computational modeling of CHD has already been explored to describe its complex flow phenomena. Although realized in other pathologies, the full potential of medicine-based evidence in CHD care remains unexplored[44].

INTERVENTIONS AND POSTOPERATIVE MANAGEMENT

AI is great at predicting the results of CHD interventions by using information from various sources, such as electronic medical records, vital signs, laboratory results, and imaging data, to obtain a clearer picture, improve system performance, and increase accuracy, which refers to multimodal data fusion. In complex cases, AI can aid surgical planning through 3D modeling and virtual simulations, enhancing precision and safety[45]. AI can predict the hemodynamic effects of interventions and optimize prosthesis design[46-48]. AI-driven early warning systems for hemodynamic deterioration are already employed for adults[49]; however, their pediatric application requires further development and validation. In the neonatal intensive care unit, AI algorithms monitor vital signs and predict clinical deterioration, which improves response times and outcomes[50]. A random forest model demonstrated high accuracy in predicting Glenn pressures > 15 mmHg following bidirectional Glenn shunt from the postoperative computed tomography, potentially reducing the need for postoperative right-heart catheterization[51]. Major improvements in predicting mortality and prolonged stay have been achieved through deep learning models, even for high-risk interventions such as the Norwood procedure for single-ventricle palliation. A deep neural network model based on the Pediatric Heart Network Single Ventricle Reconstruction trial dataset achieved 89% ± 4% accuracy and 0.95 ± 0.02 AUROC for the risk of mortality or cardiac transplantation and 85% ± 3% and 0.94 ± 0.04, respectively, for the risk of prolonged stay[52].

LONG-TERM CARE

Long-term monitoring is crucial for patients with CHD[53,54]. AI-driven wearable devices and remote monitoring offer the continuous assessment of physiological metrics, alerting clinicians to potential complications[55]. Implantable biosensors, such as cardioMEMS, provide real-time ambulatory pressure measurements[56] and are useful in managing conditions such as pulmonary hypertension and single-ventricle physiology[57-59]. AI-driven prediction of adverse events using real-time data from wearable sensors shows high accuracy in adult heart failure cohorts[60], with potential applications in long-term CHD care. Risk stratification for palliated CHDs, such as patients who have undergone Fontan procedures and are highly susceptible to ventricular dysfunction and desynchrony, is also important. In these populations, unsupervised learning from cardiac MRI data can also be used to identify high-risk cohorts[61].

CHALLENGES AND FUTURE DIRECTIONS

Despite the significant potential of AI, challenges remain. The main obstacle is that the field is still new, with most of the aforementioned work consisting mainly of preliminary tests, evaluations of model accuracy, or computer simulations. Similarly, none of the described technologies are regularly used in clinical practice, except perhaps for on-cart automatic quantification in echocardiography[21,25]. This calls for the development of more robust and clinically applicable AI tools and their clinical validation through appropriate trials.

Essentially, AI is based on data, and suitable training datasets remain one of the largest bottlenecks to such development. In CHDs, such concern is twofold.

First, data availability and “one-size-fits-all” solutions are hampered by the heterogeneity inherent in the wide spectrum of CHD. Nomenclature systems vary across practitioners. The incidence of CHDs varies geographically; however, lower- and middle-income countries have less data available. Although CHDs are not rare diseases per se, centers caring for CHDs often have relatively few patients with any given pathology. This renders accumulating a large volume of cross-sectional and longitudinal data on specific disorders and the creation of associated robust predictive models difficult[1,62].

Thus, implementing standardized nomenclature and coding systems, thorough cross-sectional and longitudinal data collection, wider adoption of electronic medical records, and increased data availability for evidence generation in CHDs is necessary[62,63]. Centralized CHD registries, such as the Fontan Outcomes Network in the United States, are the next logical step. Similar registries have also emerged in other parts of the world, such as the PartneRships in cOngeniTal hEart disease cohort in Southern Africa or the Down syndrome with CHD registry in Thailand[64,65]. These initiatives for data collection and sharing, along with cross-institutional integrations, are crucial for the further development of AI solutions for CHD.

Interestingly, this problem can also be tackled by AI-based solutions. Federated learning is a distributed machine learning framework where a common global initial model is shared by a central server to participating network sites. The model is subsequently trained with the local data from each participating site, and the resulting model weights are sent back to the central site to develop the final model. This approach is a suitable solution when data privacy concerns restrict cross-site data sharing; however, until now, the clinical application of federated learning in CHD care has been limited[66]. Similarly, data augmentation allows the expansion of the available training dataset space by making slight changes to the dataset or using deep learning techniques to generate new data points[67]. However, the clinical implications of models trained on augmented datasets in the exceedingly diverse field of CHDs remain to be explored.

Another data-related major weakness of machine learning and AI is the quality and representation of data. Training databases on disproportionately representative data trains biased models that lack discrimination for minority class instances. Wide and adequate representation is paramount to mitigate such algorithmic bias; however, obtaining such datasets is extremely challenging due to the reasons outlined above[68].

The goal of AI in CHD management is to seamlessly integrate AI tools in daily medical practices, ensuring regular updates and rigorous quality checks to maximize the potential of AI. However, numerous barriers to such integration remain, along with many unanswered questions[69]. For instance, what should be the level of autonomy granted to such AI models? What tests are necessary before that “trust” can be established? Once we grant trust in AI, who bears the responsibility for any misdiagnosis, oversight, or medical error arising from its use? These ethical and legal questions are widely discussed and reviewed and indeed go beyond the scope of this review[70,71]. The “black box” nature of various machine learning models—meaning their internal algorithms and decision-making pathways are not readily understandable to human users—remains one of the major obstacles to the clinical acceptance of such tools. Explainable AI—models that develop algorithms comprehensible to users—is a recent focus of research[72,73]. These algorithms may go a long way to secure clinician trust and allow better allocation of accountability[70]. Overall, despite the availability of guidelines and statements about the possible development of AI, regulatory frameworks for the same remain scarce and much awaited[74,75].

CONCLUSION

AI offers transformative potential for improving the diagnosis, management, and long-term care of CHD. The salient features are consolidated in Table 1. From enhancing prenatal screening through fetal echocardiography and genetic testing to revolutionizing postnatal diagnosis with advanced imaging and signal processing, AI is poised to address many critical challenges in CHD care. Compared with traditional methods, AI-driven decision support systems offer personalized medicine approaches that optimize treatment strategies and predict patient outcomes with greater accuracy. Furthermore, AI-enabled remote monitoring and wearable technologies facilitate ongoing surveillance, enabling early detection and intervention of potential complications both in the immediate postoperative period and during long-term management. However, the successful integration of AI in CHD care hinges on overcoming key challenges, including data standardization, algorithmic transparency, ethical considerations, and robust integration into existing clinical workflows. Encouraging collaborative efforts in data sharing, promoting the development of understandable AI, and ensuring responsible application can help unlock the full potential of AI to significantly improve the lives of patients with CHD.

Table 1 Role of artificial intelligence in congenital heart disease.

Topic	Specific issue	Description
Prenatal diagnosis & risk stratification	Enhanced fetal echocardiography using AI	Improved accuracy and efficiency in detecting CHD prenatally
	AI-powered genetic testing	More precise identification of genetic predispositions to CHD
Postnatal diagnosis & management	AI-driven diagnostic imaging analysis	Faster and more accurate identification of CHD subtypes and severity
	Advanced signal processing	More precise assessment of hemodynamic parameters
	AI-powered decision support systems	Personalized treatment strategies based on individual patient data
Long-term care & monitoring	AI-enabled remote monitoring	Continuous surveillance for early detection of complications
	Wearable technology	Continuous data collection and transmission for proactive management
Challenges & limitations	Need for standardized datasets	Lack of large, high-quality datasets hinders AI model development
	Development of transparent and explainable AI algorithms	Ensuring trustworthiness and accountability of AI systems
	Ethical considerations	Addressing issues of data privacy, bias, and equitable access
	Seamless integration into clinical workflows	Efficient incorporation of AI tools into existing healthcare practices

AI: Artificial intelligence; CHD: Congenital heart disease.