Congenital heart defects are relatively common, among which atrial septal defects (ASDs) are the most prevalent in adults. Without timely treatment, ASDs may lead to permanent cardiovascular damage with potentially fatal complications. However, ASDs are often diagnosed late or not at all due to the mild or asymptomatic nature of these defects, as well as the lack of overt examination findings.
A new study published in eClinicalDiscovery explores the use of artificial intelligence (AI) in enhancing the yield of conventional screening for ASDs using 12-lead electrocardiography (ECG).
Study: Deep learning-based model detects atrial septal defects from electrocardiography: a cross-sectional multicenter hospital-based study. Image Credit: Duangnapa Kanchanasakun / Shutterstock.com
ASDs increase the risk for cardiovascular complications such as atrial fibrillation (AF), heart failure, stroke, and pulmonary hypertension. Thus, the timely closure of ASDs, which can often be achieved through minimally invasive techniques, results in better life expectancy.
Typically, the diagnosis of ASDs occurs as an incidental finding or after symptoms appear later in life. Screening for the condition would require routine echocardiography, which is accurate, sensitive, and non-invasive. However, the associated cost of this technique, the need for trained personnel, and the time required prevent its practical use as a population-based screening method.
ECG is a rapid technique that requires less training and uses unsophisticated equipment, thus making it ideal for ASD screening. However, ECG changes are poorly sensitive and specific, thus causing this screening approach to often miss ASD patients and, as a result, increases their risk for adverse cardiovascular outcomes.
Current limitations in diagnostic technologies have led researchers to utilize deep learning models for the diagnosis, prognosis, and automated assessment of diseases using raw ECG data.
In the current study, researchers utilize a convoluted neural network (CNN) to analyze conventional ECG results. CNN allows for the expansive parallelized study of all filters, which is comparable to other neural network approaches like recurrent neural networks that depend on the data presented from the previous step.
What did the study show?
The study covered two continents and three hospitals, including two in Japan and one in the United States. Taken together, approximately 81,000 participants were included in the analysis, with over 671,000 ECGs used for model development.
All participants had one or more echocardiograms and were classified as being positive or negative for ASD. Any ECG from a patient with a closed ASD was excluded from the study.
Patients with ASD were significantly younger than those without, with mean ages of 41-57 years and 62-64 years, respectively.
After training the CNN-based model, researchers used the earliest ECG for each patient in the randomly assigned test set to test for ASD. This was compared to the use of either overt ECG changes caused by right bundle branch block, right atrial dilation, or any ECG abnormality.
The CNN technique successfully discriminated between positive and negative ASD cases. The area under the receiver operating curve (AUROC) was 0.85-0.90, thus demonstrating that this screening strategy effectively distinguished between patients with and without ASD in up to 90% of cases.
Subgroup analysis was performed using various patient characteristics like age, sex, body mass index (BMI), and the presence of AF or any ECG abnormality. The results confirmed the excellent capability of the deep learning algorithm to detect ASD.
The model performed well with mean pulmonary artery pressure (PAP) below or above 20 mm Hg. The highest discrimination was shown with severe ASD, in which the AUROC reduced to 0.65 for ASD smaller than 10 mm and 0.76 for catheterization readings Qs/Qp less than 1.5.
The CNN model also identified ASD where there were indications for closure, as reflected by Qp/Qs values exceeding 1.5, with AUROC of 0.91. These results were comparable, irrespective of BMI from less than 18.5 to 25 or higher. The positive predictive value was 14%, which is comparable or better than the 13.6% with ECG abnormalities, at the same sensitivity of 79%.
The simulated model resulted in a significantly improved sensitivity of screening, from about 80% based on ECG abnormalities observable by clinicians to almost 94% with the use of the CNN-based model. The specificity in both cases was 33.6%.
The consistent performance of the model across patients from different sociocultural and ethnic settings, as well as geographic locations, indicates its high generalizability.
This study showed that a neural network-based DL algorithm using 12-lead ECG data can detect ASD excellently with good generalization. The model can be used to improve ASD screening, where symptoms and laboratory findings are subtle.”
What are the implications?
DL models can extract unrecognized information from ECGs and incorporate minute and bias-free features that are often missed by the human eye, ultimately enabling the construction of more accurate algorithms.”
The CNN model successfully achieved sensitive detection of ASD-suggestive ECG patterns without further reducing the specificity as compared to conventional ECG-based screening. In addition to its discriminative power, this tool can maintain a consistent specificity and sensitivity across institutions, thus establishing its generalizability and excellent utility for widespread use in ASD screening in patients with subtle signs and laboratory findings.
Nevertheless, future prospective trials on the general population are needed to ensure the utility of this approach in asymptomatic individuals.
The application of AI to the increasingly diverse and extensive collection of ECGs from congenital heart disease patients at various stages of disease and at different periods of time will likely help detect these conditions earlier, thereby allowing for more effective intervention.