EP09.12: Use of artificial intelligence for automatic antenatal classification of congenital heart diseases

To analyse the diagnostic performance of artificial intelligence for antenatal classification of congenital heart diseases. Cross-sectional study on obstetric ultrasound images in health centers involved in the program for the diagnosis, training, logistics, education, organisation, and follow-up of patients with congenital heart diseases (DELFOS) from 2019 to 2023. A protocol was designed in conjunction with specialists from health centers, including standardisation factors such as gestational trimester, types of cuts (four chambers, three vessels trachea, right ventricular outflow tract, left ventricular outflow tract), and additional information necessary for a correct diagnosis. A model based on a ResNet (He, 2016) was developed, which provided a first classification regarding the presence or absence of heart diseases, and finally, a Swin Transformer model (Liu, 2021) was used, which creates hierarchical feature maps by merging images at deeper layers. The diagnostic accuracy of artificial intelligence was estimated using prenatal ultrasound diagnosis confirmed with postnatal echocardiography as the gold standard. 583 images of congenital heart diseases (CHD) and 572 images of healthy fetuses were included. The model receives an ultrasound image as input, then performs an analysis at different scales to reach the final result determining the probability of having a CHD. It was evidenced that the model has a diagnostic performance with a sensitivity of 80%, specificity of 60%, positive predictive value of 66.7%, negative predictive value of 75%, false positive rate of 40%, and false negative rate of 20%. A high precision in the diagnostic performance of artificial intelligence for the identification of CHD was demonstrated, using prenatal ultrasound diagnosis confirmed with postnatal echocardiography as the gold standard. This highlights the effectiveness of the AI approach in detailed and accurate interpretation of images.