JMIR Publications

ABSTRACT

Background:

Sub-Saharan Africa (SSA) has the highest global burden of under-five child mortality, with congenital heart disease (CHD) being a major contributor. Despite advancements in high-income countries, CHD-related mortality remains unchanged in SSA due to limited diagnostic capacity and centralized healthcare. While pulse oximetry supports early detection, confirmation of diagnosis often relies on echocardiography, a procedure hindered by a shortage of specialized personnel. Artificial intelligence (AI) offers a promising solution to address this diagnostic gap.

Objective:

This study aims to develop an AI-assisted echocardiography system that will enable non-expert operators such as nurses midwifes and medical doctors to perform basic cardiac ultrasound sweeps on neonates suspected of CHD and extract accurate cardiac images that can be sent to a remote paediatric cardiologist for interpretation

Methods:

The study will follow a two-phase approach to develop a deep learning model for real-time cardiac view detection in neonatal echocardiography, using data from St Padre Pio Hospital in Cameroon and the Red Cross War Memorial Children’s Hospital in South Africa ensuring demographic diversity. Phase one will pretrain the model on retrospective data from ~500 neonates (0–28 days old). Phase two will fine-tune it using prospective data from 1,000 neonates, which includes background elements absent in the retrospective set, enabling adaptation to local clinical environments. The datasets will include short and continuous echocardiographic video clips covering ten standard cardiac views, as defined by the American Society of Echocardiography. The model architecture will leverage convolutional neural networks (CNNs) and convolutional Long Short-Term Memory (convLSTM) layers, inspired by the interleaved visual memory framework, which combines fast and slow feature extractors through a shared temporal memory mechanism. Video preprocessing, annotation with predefined cardiac view codes using Labelbox, and training with TensorFlow and PyTorch will be conducted. Reinforcement learning will guide the dynamic use of feature extractors during training. Iterative refinement, supported by clinical input, will ensure the model effectively distinguishes correct from incorrect views in real-time, enhancing usability in resource-limited settings.

Results:

Retrospective data collection for the project began in September 2024, and since then, data from 308 babies have been collected and labelled. In parallel, the initial model framework has been developed and training initiated using a small portion of the labelled data. The project is currently at the intensive execution phase with all objectives running in parallel and final results expected within 14 months.

Conclusions:

The AI-assisted echocardiography model that will be developed in this project holds promise for improving early CHD diagnosis and care in SSA and other low resource settings.