CNN-robust transformer for diastolic dysfunction detection using combined heart sounds and electrocardiogram signals

Abstract Left ventricular diastolic dysfunction (LVDD) is a critical indicator of cardiovascular health, strongly associated with adverse outcomes such as increased mortality and hospitalization due to heart failure. Non-invasive approaches, including heart sounds and electrocardiograms (ECG), have demonstrated potential for LVDD detection. The combined use of multimodal signals can provide richer physiological and pathological information, but it may also introduce additional noise, potentially affecting diagnostic accuracy. Transformer-based models have recently shown strong capabilities in capturing long-range dependencies and integrating information across modalities, yet they remain susceptible to noise. To address this challenge, we propose a Robust Transformer (RT) module that enhances the model’s resilience by dynamically adjusting the value matrix in each attention head, thereby suppressing noise-sensitive responses and stabilizing attention aggregation. By integrating the RT module with convolutional neural networks (CNNs), which are effective at extracting local features, we develop a lightweight and robust framework for LVDD identification that leverages the complementary strengths of multimodal signals while mitigating their potential noise. Finally, by jointly leveraging raw signal segments, physiological features, and extracted representations, the proposed model achieved strong performance in diastolic function identification, highlighting its effectiveness and potential for non-invasive LVDD screening.