Enhanced detection of aortic stenosis using a multimodal deep learning approach combining ECG and chest X-ray data

Abstract

Listen

Translate article

Abstract Background/Introduction The early detection of aortic stenosis (AS) is crucial due to its increasing prevalence in the aging population. In Japan, approximately 2.84 million people over the age of 60 are estimated to have AS, with around 560,000 requiring surgical intervention. However, the annual number of treatments, including transcatheter aortic valve implantation and surgical aortic valve replacement, is limited to about 20,000 cases, far fewer than needed. Early diagnosis and treatment are essential for a better prognosis, whereas delays result in poor outcomes. Although transthoracic echocardiography (TTE) is the standard for diagnosing AS, it is often performed on symptomatic patients, and the time required to gain proficiency in TTE makes it unsuitable for widespread screening. In contrast, ECGs and chest X-rays are more accessible and commonly used inhealth checkups. Recent advances in AI have shown potential in predicting valvular diseases from ECGs and AS from chest X-rays, but using these methods individually has limitations. We hypothesized that combining ECG and chest X-ray data could enhance the accuracy of AS predictions through a deep learning model. Methods We utilized data from patients who underwent ECG, chest X-rays, and echocardiography at the Department of Cardiology, Kobe University Hospital, between January 1, 2012, and December 31, 2022. Patients with mild or more severe AS were considered positive cases. Mild AS was defined as Vmax ≥ 3.0 m/s, mPG ≥ 20 mmHg, and AVA ≤ 1.5 cm².The dataset, consisting of 23,886 patients with 1,442 positive cases, was divided into training (14,438 patients), validation (4,673 patients), and test (4,775 patients) datasets. The ECG model was developed using ResNet and Transformer architectures, while the chest X-ray model used EfficientNet. These models were combined using cooperative learning loss. The model's performance was evaluated using ROC curves and AUC, accuracy, precision, recall, F1-score for binary classification of AS. Results The cooperative learning model achieved a test accuracy of 0.879, a test AUC PR of 0.300, a test AUC ROC of 0.822, a test recall of 0.493, a test specificity of 0.904, and a test precision of 0.247 in the binary classification of AS. Conclusion(s) The multimodal deep learning model that combines ECG and chest X-ray data proved effective in diagnosing aortic stenosis, demonstrating its potential as a valuable tool for early detection and improved patient outcomes.