SAGE‐NET—Reconceiving Edge Intelligence Using Attention‐Based Deep Learning Framework for Effective Classification of Electrocardiogram (ECG) in WSN‐IoT Environment

With the continuous enhancement of heart disease monitoring systems, awareness of cardiac diseases has increasingly captured public attention, leading to an exponential rise in demand for intelligent health status monitoring systems. Currently, these cardiac status monitoring systems largely depend on heavy equipment, which is costly and lacks flexibility and scalability. Recently, wireless sensor network–enabled Internet of things (WSN‐IoT) has played a crucial role in integrating intelligent device capabilities into diagnosis and clinical treatment by replacing traditional equipment. Additionally, edge computing is a novel paradigm that facilitates cognitive and analytical applications in WSN‐IoT devices, enhancing monitoring and prognosis. However, these intelligent devices still suffer from limited lifetime and resources, making it essential to optimize resource usage without compromising monitoring performance. By applying an intelligent approach, resource optimization can be achieved, resulting in superior classification performance. This paper proposes a novel resource‐constrained attention map using gated recurrent neural networks (GRNNs) for deployment on edge‐supported WSN devices, which classifies electrocardiogram (ECG) signals without compromising optimal resource usage. In this model, attention layers embedded in the GRNN are deployed on edge devices to enhance intelligence for effective ECG signal classification. Extensive experimentation was conducted with 50 patients using real‐time evaluation boards with embedded microcontrollers connected to ECG devices and NVIDIA Jetson Nano boards as edge computing devices. To highlight the strengths of the proposed framework, efficiency evaluations were made with other deep learning (DL) approaches for edge devices. Studies indicate that the proposed model—relied on edge‐enabled WSN devices—achieved better computational cost and classification outcomes (99% accuracy) compared to state‐of‐the‐art learning methods and was 23% less costly than other learning methods.