AI-facilitated Echocardiography for Early Detection of Acyanotic Congenital Heart Disease in Infants - AI-facilitated Echocardiography for Early Detection of Acyanotic Congenital Heart Disease in Infants Abstract The overarching aim of this study is to explore the development of artificial intelligence (AI) models for the automated detection of the five most common acyanotic congenital heart diseases (CHDs)—atrial septal defect (ASD), ventricular septal defect (VSD), atrioventricular septal defect (AVSD), coarctation of the aorta (COA), and patent ductus arteriosus (PDA)— in infants aged 0–2 years using postnatal echocardiography. Building on the foundation of our CHD AI project, which successfully automated the identification of standard cardiac views from echocardiographic videos, this sister project (CHD AI-2) will leverage existing datasets from South Africa and Cameroon to extend these capabilities towards the detection of specific acyanotic CHDs that frequently remain undiagnosed in early infancy. A central goal of CHD AI-2 is to develop AI models specifically designed for use by noncardiology expert health workers, such as nurses, midwives, general practitioners, and sonographers, working in low-resource peripheral health facilities where most births occur but where cardiologists and specialized diagnostic services are unavailable. Using convolutional neural networks, the study will explore the development of AI models that can detect the target CHDs through a dual approach: identification of anatomical landmark features characteristic of each target CHD and detection of the specific lesions that define the condition. If successful, this project will complement the AI models developed for cardiac view extraction in the initial CHD AI project by enabling the identification of specific acyanotic CHDs from these automatically extracted cardiac views. In clinical practice, during a nonexpert ultrasound examination, the combined AI models will extract appropriate cardiac views and detect features suggestive of specific acyanotic CHDs, which can then be transmitted to remote expert cardiologists for confirmatory review and diagnosis. Ultimately, working together, the AI models from the CHD AI and CHD AI-2 projects will help bridge the diagnostic gap in CHD detection, facilitating timely referral and intervention and contributing to the reduction of preventable infant mortality in the region. Beyond its direct application in sub-Saharan Africa, the project has potential benefit to the United States through the generation of a valuable, expert-annotated dataset of infant echocardiographic videos, including images acquired by non-expert health workers under real-world, resource-constrained conditions. These data and the resulting AI models will provide an important research resource for developing and evaluating robust, operatorindependent cardiac imaging technologies. The knowledge generated may also inform future approaches to CHD screening in rural, remote, and resource-constrained US settings where access to pediatric cardiology expertise may be limited. Thus, the project has the potential to contribute not only to improved early CHD detection in Africa, but also to US research and innovation in scalable AI-enabled pediatric cardiac screening.
