Related Experiment Videos
HPA-Net: a lightweight hierarchical progressive fusion network for cardiac disease detection based on ECG-PCG signals
Tong Xiao1,2, Lufeng Che1,2
1College of Information Science and Electronic Engineering, Zhejiang University, No. 38 Zheda Road, Xihu District, Hangzhou 310027, People's Republic of China.
Insights
A new deep learning model, HPA-Net, efficiently diagnoses cardiovascular disease using electrocardiogram (ECG) and phonocardiogram (PCG) signals. This lightweight, dual-modal framework offers high accuracy with low computational cost for real-time monitoring.
Area of Science:
- Biomedical Engineering
- Artificial Intelligence in Healthcare
- Cardiovascular Diagnostics
Background:
- Cardiovascular disease (CVD) is a major global health concern requiring rapid diagnostic tools.
- Current diagnostic methods often fail to fully leverage the heart's bioelectromechanical properties.
- There is a need for efficient, dual-modal diagnostic frameworks for resource-limited settings.
Purpose of the Study:
- To introduce the Hierarchical Progressive Attention Fusion Network (HPA-Net), a lightweight deep learning model for CVD diagnosis.
- To enable efficient and reliable diagnostic performance on devices with constrained resources.
- To improve the utilization of combined electrocardiogram (ECG) and phonocardiogram (PCG) signals.
Main Methods:
- Developed a compact deep learning architecture (HPA-Net) with 0.21 million parameters.
- Employed a multi-scale convolutional attention encoder for spatiotemporal feature extraction from ECG and PCG signals.
- Utilized a multi-level feature integration strategy and Fast Adaptive Multitask Optimization (FAMO) for robust training.
Main Results:
- Achieved high accuracy (97.59%), sensitivity (98.42%), and specificity (95.53%) on the PhysioNet/CinC 2016 dataset.
- Demonstrated strong generalization with 97% accuracy on unimodal datasets.
- Exhibited low inference latency (41.20 ms) and a Real-Time Factor (RTF) of 0.008.
Conclusions:
- HPA-Net efficiently processes complex spatiotemporal data with minimal computational requirements.
- The model's high accuracy and low latency support deployment on edge hardware for real-time cardiovascular monitoring.
- HPA-Net presents a promising solution for early CVD detection in clinical and home-care settings.
Abstract:
Objective.Cardiovascular disease remains a leading global health threat, creating a strong clinical need for convenient and rapid diagnostic methods. However, current single-modality and conventional fusion methods often fail to adequately utilize the intrinsic bioelectromechanical coupling characteristics of the heart. To address these limitations, we propose the hierarchical progressive attention fusion network (HPA-Net), a lightweight, dual-modal framework designed for efficient and reliable diagnostic performance on resource-constrained devices.Approach.HPA-Net is a compactdeep learning architecture comprising only 0.21 million parameters and 0.362G FLOPs. It utilizes a multi-scale convolutional attention encoder to extract spatiotemporal representations from electrocardiogram and phonocardiogram signals. A multi-level strategy was employed to integrate the features across multiple processing stages. Furthermore, the fast adaptive multitask optimization technique was implemented to strengthen multi-branch learning and mitigate the issue of modality dominance during the training process.Main results.Rigorous validation on the PhysioNet/CinC 2016 subset-A dataset demonstrated that HPA-Net achieved an accuracy of 97.59 ± 0.72%, sensitivity of 98.42 ± 0.46%, and specificity of 95.53 ± 1.87%, achieving highly competitive accuracy with drastically reduced computational overhead. Despite its compact size, the model exhibited strong generalization, reaching 97% accuracy on the independent unimodal datasets. Technical evaluation showed an inference latency of 41.20 ms and a real-time factor of 0.008.Significance.The proposed architecture efficiently acquires complex spatiotemporal representations while maintaining a minimal computational demand. Its high accuracy combined with low inference latency confirms the feasibility of HPA-Net for seamless deployment on edge hardware. providing a promising solution for real-time cardiovascular monitoring and early disease detection in practical clinical or home care scenarios.
