Artificial intelligence in the prenatal detection of heart diseases

Welcome to Health Connect, the podcast for health professionals where we will share the latest news and information on science and technology in the medical industry. In this episode we will talk about the prenatal detection of heart diseases using artificial intelligence.

Did you know that congenital heart disease can be asymptomatic in fetal life but causes a high mortality after birth?1

Today we will present two studies that exemplify the use of artificial intelligence, or AI in cardiology, particularly in the detection of congenital heart disease using imaging approaches during the prenatal period.

As you may already know, the current recommendation worldwide in pregnant women is to perform a fetal screening ultrasound during the second trimester, which helps the diagnosis of congenital heart disease, or CHD, which is the most common birth defect.1

Contrary to the postnatal diagnosis, fetal diagnosis is critical to improve neonatal outcomes and help interventional planning of in utero therapies. But to reach such outcomes, it is key to distinguish normal fetal hearts from CHD that require referral to a fetal cardiologist.1

In practice, however, there is a very low sensitivity, often as low as 30 percent, that limits the treatment options and worsens the postnatal outcomes. Furthermore, a fetal screening ultrasound generally includes thousands of image frames spanning multiple structures, so the diagnostic frames of interest for CHD may be only a handful that are thus easily missed.1

In 2021, Rima Arnaout and colleagues, from the Division of Cardiology in the University of California, published in the Nature Medicine journal, a research that aimed to test if AI could be used to improve fetal CHD detection.1

The challenging goal of the researchers was to use neural networks to, first identify the five diagnostic-quality cardiac views from all images in a fetal ultrasound, second, to use these views to provide classification of normal heart versus CHD, and third, to calculate the biometric measurements for each cardiac chamber.1

For this, the researchers used up to 107 823 multimodal images from 1 326 studies, together with experts in fetal cardiology to train the neural network model, followed by five different datasets for testing. The images came from either fetal echocardiograms or fetal surveys, which are second trimester obstetric anatomy ultrasounds.1

So, first, the researchers wanted to identify the five views of the heart recommended for fetal CHD screening out of the whole population of images obtained during a fetal scan. Thus, they trained a convolutional neural network view classifier to pick the five screening views from fetal ultrasounds, for which any image that was not one of the five guideline-recommended views was classified as non-target, for example, head, foot, placenta. Notably, this view classifier was able to successfully select these images with a sensitivity and specificity of up to 96 and 92 percent, respectively.1

Next, the authors aimed to classify the views into normal hearts or complex CHD. For this, they trained the same neural network architecture used before with a dataset that contained both normal and abnormal images. After testing the model in five different datasets, the sensitivity of the model was 95 percent, and the specificity 96 percent, with a positive predictive value of 20 percent and a negative predictive value of 100 percent in distinguishing normal from abnormal hearts. Interestingly, the abdomen view was the least useful for the diagnosis of CHD.1

In addition, the authors were further interested in detecting CHD in ultrasounds of twins and other multiple pregnancies, since it is known that multifetal pregnancies have a higher risk of CHD than the general population. The results of the model on ten sets of twins, including those with tetralogy of Fallot and hypoplastic left heart syndrome, showed a sensitivity and specificity of 100 and 72 percent at predicting views and diagnoses, respectively.1

Finally, the researchers also wanted to determine biometric measurements, like the cardiothoracic ratio, cardiac axis and fractional area change, because these measurements are known to help in the screening and diagnosis of fetal CHD. To achieve this, they trained a model to find cardiothoracic structures in axial four chamber images and used these segmented structures to calculate the biometric measurements, showing that the predicted measurements have good agreement with the clinical findings.1

The relevance of this study relies on the fact that, although CHD is the most common birth defect, it is still relatively rare and one of the most difficult diagnostic challenges in ultrasound. And despite this low prevalence, the authors were able to find diagnostic signals that allowed accurate predictions of CHD, which is key to translating this work toward the real world1. Thus, to the best of the authors’ knowledge, this is the first use of deep learning to assess community-level sensitivity and specificity on a global diagnostic challenge in a test set of real fetal screening ultrasounds with a CHD prevalence similar to the general population.1

CAPSULA

The methods of choice for the detection of congenital heart disease are the second trimester ultrasonography and fetal echocardiography, but they are still far from being optimal despite the efforts in standardization of the technique and improvements in technology. In fact, there are considerable variations within geographical areas in regard to the diagnosis of CHD, ranging in severe cases from 13 to 87 percent.2

A study examined why congenital heart diseases are missed during prenatal diagnosis and found that in the majority of cases, either the sonographic plane was not correctly obtained, or, even if the defect is clearly evident on screen, the operator fails to recognize it.2 Other reasons for the low diagnosis rates are the inadequate acquisition of diagnostic-quality images due to poor acoustic windows, fetal motion or the small size of the fetal heart.1

To try to solve these problems, ultrasound equipment manufacturers have introduced AI-driven obstetric ultrasound products to the market. These softwares aim to enhance sonographers’ skills in identifying ultrasonographic structures and performing automatic biometric measurements, which makes the workflow faster. These approaches have the potential to greatly improve the prenatal detection rate of CHD.2

END OF CAPSULA

Welcome again. In the previous section, we talked about how artificial intelligence could help in the detection and diagnosis of CHD from ultrasound images using a convolutional neural network framework. Now we will talk about how other artificial intelligence approaches can also be useful to diagnose these defects during the prenatal period.

Here, we will focus on the research of Truong and colleagues, from the Christ Hospital Health Network, who published in The International Journal of Cardiovascular Imaging the results of their retrospective study that aimed to investigate the impact of their artificial intelligence framework in the prediction of either the presence or the absence of congenital heart disease using fetal echocardiography. Contrary to the study we discussed before the capsule, this work did not use convolutional neural networks, but a machine learning approach called Random Forest.

To do this, they used a database of 3 910 singleton fetuses of 22 weeks of gestational age at the time of the echocardiography. The database consisted of a comprehensive fetal echocardiogram that included 2D cardiac chamber quantification, valvar assessments, assessment of great vessel morphology, and Doppler-derived blood flow interrogation.3

For each patient, the database contained 25 demographic and clinical variables, including echocardiographic dimensions, like cardiothoracic ratio, cardiac axis, left and right atrium diameter, and mitral valvar annulus diameter. The velocity of blood flow across cardiac valves and vessels derived from Doppler was recorded as well.3

In this population, the proportion of CHD was 14.1 percent as confirmed by post-natal echocardiograms, with the most common defects being atrioventricular and ventricular septal defects, and hypoplastic left heart syndrome.3

The data with the 25 variables was used to train and test their Random Forest-based model. Their framework provided a sensitivity of 0.85, a specificity of 0.88, a positive predictive value of 0.55 and a negative predictive value of 0.97 to detect CHD. Thus proving that the model can effectively detect patients with CHD with accuracy and reliability.3

In addition, a detailed feature importance analysis showed that using less than 6 out of all the variables of the echocardiogram does not have a satisfactory performance when classifying into CHD or no CHD. Instead, there are 6 top features, including cardiac axis, peak velocity of blood flow across the pulmonic valve, cardiothoracic ratio, pulmonary valvar annulus diameter, right ventricular end-diastolic diameter, and aortic valvar annulus diameter, which are essential and add more predictive values to the model in detecting patients with CHD.3

Despite the promising results, the study has some limitations that have to be taken into account, for instance, this was a single-center investigation, and thus selection bias by population-specific to the location of the center may have been introduced. Also, the study did not include an assessment of raw images specific for analysis. Finally, while the study demonstrated good sensitivity, it did not specify each CHD lesion.3

In conclusion, the two works presented today demonstrate that combining guideline-recommended imaging with ensemble learning models, like convolutional neural networks, could significantly improve detection of fetal CHD.1 Moreover, Random Forest-based frameworks can provide increased sensitivity in prenatal CHD screening with very good performance.3

The incorporation of machine learning algorithms into fetal echocardiography may further standardize the assessment for CHD, which may be of particular benefit to the echocardiographer with limited experience.3 Finally, testing and refining this ensemble learning models in larger populations will help to democratize the expertise of fetal cardiology experts to providers and patients worldwide and to apply similar techniques to other diagnostic challenges in medical imaging.1

Thanks for joining us on this episode of Health Connect. Don’t miss out on our next episode. Discover more medical news and content on Viatris Connect.



Referências:

  1. 1. Arnaout R, Curran L, Zhao Y, Levine JC, Chinn E, Moon-Grady AJ. An ensemble of neural networks provides expert-level prenatal detection of complex congenital heart disease. Nat Med. [Internet]. 2021. (Accessed on May 31, 2022); 27(5):882–891. Available at: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC8380434/ Author manuscript; available in PMC 2021 August 2.
  2. 2. Rizzo G, Pietrolucci ME, Capponi A, Mappa I. Exploring the role of artificial intelligence in the study of fetal heart. Int J Cardiovasc Imaging. [Internet]. 2021. (Accessed on May 31, 2022);38:1017–1019. Available at: https://doi.org/10.1007/s10554-022-02588-x
  3. 3. Truong VT, Nguyen BP, Nguyen-Vo TH. Application of machine learning in screening for congenital heart diseases using fetal echocardiography. Int J Cardiovasc Imaging. [Internet]. 2021. (Accessed on May 31, 2022);38:1007–1015. Available at: 3. https://doi.org/10.1007/s10554-022-02566-3

Os links para todos os sites de terceiros são oferecidos como um serviço aos nossos visitantes e não implicam endosso, indicação ou recomendação do Health Connect. Os artigos vinculados são fornecidos apenas para fins informativos e não visam implicar uma atribuição pelo autor e/ou editor. O Health Connect se isenta de qualquer responsabilidade pelo conteúdo ou pelos serviços de outros sites. Recomendamos que você analise as políticas e condições de todos os sites que escolher acessar.

NON-2022-12477
NON-2023-2512