Congenital heart disease (CHD) is a common and serious congenital anomaly worldwide, representing the primary contributor to infant mortality associated with birth defects (1–3). Globally, the prevalence of CHD at birth currently is 9.4‰ (2), translating to approximately one new case every three minutes. Therefore, achieving a timely and precise prenatal diagnosis of CHD is imperative, laying the foundation for informed perinatal decision-making and significantly impacting rates of morbidity and mortality (4, 5). Furthermore, prenatal genetic diagnosis of CHD holds profound implications for clinical management, prognosis, genetic counseling, and preventive measures for subsequent pregnancies.

Despite the considerable advances in prenatal diagnostics and management strategies for CHD, it remains the leading cause of neonatal mortality (6, 7). The existing gap in prenatal diagnosis can be attributed to two primary factors. Firstly, the accuracy of prenatal ultrasound diagnostics varies considerably, owing to challenges in obtaining standardized fetal cardiac ultrasound images and interpreting results effectively (8–10). Secondly, the multidimensional nature of prenatal data poses challenges in achieving a comprehensive diagnosis (11, 12).

Obtaining standard views for the prenatal diagnosis of CHD is a complex and uncertain task. This is due to the small size, the fast beating of the fetal heart, the fetus rotating in utero, and the maturity of its structure with gestational age (13). To achieve accurate diagnostic views, the operator must thoroughly understand cardiac anatomy. Additionally, they need a strong spatial imagination and must adjust spatial relationships based on fetal position and gestational age. Therefore, this task poses a challenge for most prenatal screening physicians.

Moreover, diagnosing CHD involves processing highly complex multidimensional information, including organ structure, gene ontology, maternal disease, and clinical pathways. These pieces of information are related to each other, affect each other, and jointly impact prognosis. The weight of each information on prognosis is also different. Furthermore, there are more than 200 sub-types of CHD with significant differences in outcomes and more than 400 known causative genes for CHD (14). At the same time, many potential causative genes and many kinds of maternal diseases are related to fetal cardiac defects, such as autoimmune diseases, gestational diabetes, and infectious diseases. Changes in any variable can lead to changes in clinical decision-making. For example, if a fetus presents with a straightforward ventricular septal defect (VSD), and without additional anomalies, the prognosis tends to be favorable. Contrastingly, when a VSD coincides with chromosomal abnormalities, like the deletion of chromosome 5’s short arm leading to “cat’s cry syndrome,” the prognosis becomes grim. In fact, only a minority of these children survive into adulthood, often exhibiting evident intellectual disabilities (15).

The application of AI in aiding medical diagnosis is growing as it offers fast processing, improved accessibility, and enhanced work efficiency. This has led to a greater demand for AI in clinical settings. Recent advancements in AI techniques have led to numerous studies on CHD, aimed at optimizing the scanning process, improving image quality, and enhancing screening and diagnostic capabilities (16–18). There are many data sources and clinical decision-making processes in these applications. Moreover, numerous AI-based methods have been proposed for analysis specific to different data types. Although several review articles have been written on the application of AI in CHD, these articles focused on either the whole field of cardiology (19–22) or on other single-imaging modalities such as ultrasound images (23, 24). Also some specific tasks like image segmentation have been summarized (25), there is a lack of overall review analysis of AI for CHD. This article addresses the absence of a comprehensive review analysis of AI for CHD by reviewing existing AI-powered applications. It aims to assist beginners and non-specialists in gaining a better understanding of this relatively new technique while also promoting further investigations and applications in the field.

In this review, we introduce the data sources and clinical decision-making processes in CHD that can be addressed with AI methods and discuss the different applications of AI for CHD. We start by presenting AI concepts and illustrating the prenatal diagnosis of CHD in imaging that AI algorithms can currently perform. Then, we present how CHD genetic prenatal diagnosis knowledge can be introduced into AI applications through analytical modeling. Next, we demonstrate the AI applications of underlying factors associated with the development of CHD. Finally, we discuss the remaining challenges to the widespread use of AI in CHD.

For over two decades, AI has been integral to advancements in the medical field, playing a significant role in developing computer-aided diagnosis, treatment, and decision-making that assist clinical physicians (26). Machine learning is a branch of artificial intelligence that gradually improves statistical methods with increased data to obtain the best model and predict unknown conditions. The basis of machine learning lies in big data, and its abundant resources come from a vast database accumulated in routine clinical practice. Machine learning algorithms require training data to acquire knowledge of the parameters involved in a particular task. Additionally, validation data is often used to optimize these parameters. The accuracy of the task is then assessed using test data that was not previously seen during the learning process. Three types of machine learning methods exist: supervised, unsupervised, and semi/weakly-supervised learning (27). Labeled training data establish the relationship between input and output in supervised learning. For example, the input often consists of images or videos, clinical information (such as maternal factors including age, occupation, medication exposure and mental stress), and gene data. Human annotations can determine the ground truth for supervision, CHD fetal autopsy or postnatal validation, and clinical CHD phenotype. In unsupervised learning, the algorithm clusters unlabeled data to find inherent similarities, and only input data is used without additional annotations. Therefore, a larger amount of data is typically required. Semi/weakly learning aims to achieve the same results as supervised learning while using the minimum amount of labeled data possible. Human annotation is usually a challenging and time-consuming task that can also be influenced by observer dependence. Hence, minimizing the need for annotation can reduce annotation costs and improve the model’s potential for generalization.

Since 2015, AI-powered prenatal diagnosis of CHD has witnessed substantial growth, with a focus on improving data acquisition and assisting tools to enhance screening and diagnostic capabilities. Unlike static organs and adult hearts, prenatal diagnosis of CHD presents specific challenges such as data collection through the maternal uterus, rapid pulsation, small size, and organ maturation during gestational age. The application of AI technologies provides clinicians with convenient tools. For instance, semantic segmentation aids in identifying cardiac structural anomalies, facilitating easier diagnosis for medical professionals. Image classification helps doctors quickly locate the desired fetal cardiac views, while key-point detection assists in obtaining more accurate structural measurement data.

The primary tasks involved in AI-based prenatal CHD diagnosis encompass CHD imaging diagnosis, CHD-related genetic diagnosis, and analysis of prenatal risk factors associated with CHD (Figure 1). In the following, we will review the major progress made in these areas.

AI methods necessitate four research tasks for prenatal diagnosis of CHD in imaging: acquisition and reconstruction, quality control, assisted analysis tools, and screening and diagnosis. Figure 2 illustrates the interrelationship among these tasks.

Nearly 20% of fetal cardiovascular diseases are associated with chromosomal abnormalities (42–44). Clarifying the genetic causes has a significant impact on genetic counselling and fertility guidance for the affected couples. Nonetheless, research on the genetic pathogenesis of CHD is limited by the number of family samples and the depth of genetic testing. According to the PubMed database, fewer than 20 non-case reports on whole-exome sequencing for CHD were published before December 2022, with a total sample size of fewer than 1,000 cases. Only 9 studies had a sample size greater than 30; the maximum sample size was only 235 cases (45, 46). Only a few studies using AI-based tasks with CHD genetic test data have been found. Radhakrishna et al. (47) employed logistic regression analysis to select candidate markers for VSD prediction and explored the mechanism of CHD. Sun et al. (48) evaluated the association of maternal genes with the risk of CHD with 569 eligible cases, but due to the limited sample size, the specific subtypes could not be examined. Li et al. (49) suggested the possibility of discerning the nonlinear relationship between genes and diseases from graph convolutional neural networks by considering both the network topology and multiple sources of information on diseases and genes.

A strong association exists between the incidence of CHD and exposure to risk factors, such as maternal illness, medication intake, and mental stress during early pregnancy (50–52). Nevertheless, the underlying mechanism of CHD appears to be multifactorial. Therefore, identifying potential risk factors for CHD through AI is crucial to reduce its incidence. Yang et al. (53) investigated the association between maternal exposure to air pollution and CHD. They used a logistic regression model and demonstrated that maternal exposure to any air pollutants during the first trimester is associated with increased odds of CHD. Wern et al. (54) calculated odds ratios (ORs) by comparing the live birth rates of babies born to mothers with and without diabetes. It also shows the risk of CHD for the fetuses of pregnant women with diabetes was five times greater than the fetuses of pregnant women without diabetes. According to Wurst et al. (55), the use of paroxetine and other psychotropic drugs during pregnancy increases the risk of CHD in population studies and clinical randomized trials. Karatza et al. (56) collected data on the maternal and infant characteristics of fetuses diagnosed with CHD and found that the risk of CHD increased by 2.7 times when pregnant women were exposed to smoking during pregnancy and that the risk increased with increasing smoking dose. However, more than 300 parameters can be considered fetal factors, and the teratogenic factor prioritization problem in CHD remains elusive due to the large search space and insufficient training datasets.

Prenatal Diagnosis of CHD with AI faces challenges similar to other medical AI applications. Its opaque decision-making process raises interpretability concerns crucial for clinical decision-making. Addressing ethical and legal issues, such as privacy and discrimination, requires clear guidelines. Implementing AI systems demands significant resources and may exacerbate healthcare disparities. Continuous adaptation is essential to align AI models with evolving medical knowledge. Collaborative efforts among healthcare providers, researchers, policymakers, and ethicists are crucial for responsible and ethical AI use in enhancing prenatal care and CHD diagnosis.

Additionally, current AI studies in CHD primarily replicate existing methodologies, such as classification, segmentation, and detection, relying heavily on expert annotations. They lack exploration and optimization of decision-making pathways. These studies can be summarized into three main issues: volatility, insufficiency, and independence (VII).

CHD has a high incidence and includes a wide spectrum of diseases, as well as a wide range of outcomes. This involves multi-source and multidimensional diagnostic information, introducing both challenges and opportunities to AI analysis of CHD. Current AI research for prenatal CHD diagnosis implies that it can improve the workflow of CHD screening and diagnosis. It can also increase the confidence of prenatal ultrasound diagnosticians and improve the effectiveness of prenatal CHD screening. However, the previous AI paradigm cannot meet the current clinical needs of CHD due to VII problems.

Addressing the challenges in prenatal CHD diagnosis, AI techniques can be utilized to extract deeper-level data features and integrate them into data annotations to improve data re-usability. Additional dimensions, such as sensor location and physician attention data, can be incorporated to enrich the dataset. Furthermore, optimizing clinical decision-making with AI algorithms can enhance efficiency and accuracy. As a result, AI and CHD form a bidirectional interaction cycle, requiring efforts on both the clinical and AI sides. AI drives clinical development in CHD, and CHD drives AI.