Efficient Cardiovascular Disease Diagnosis System for an Wearable Device based on Multi Stage BCResNet

Cardiovascular diseases (CVDs) remain a leading cause of global mortality, necessitating continuous monitoring, particularly through electrocardiogram (ECG) signals. Although current wearable devices detect abnormalities, they often fail to precisely identify specific diseases, leading to diagnostic delays. This paper proposes an innovative multi-stage approach employing an on-device deep neural network architecture. Our method identifies abnormal signals and classifies them into specific disease categories. Using a simple binary classifier in an iterative approach, we form a multi-label classifier that dynamically loads tailored weight parameters, minimizing resource use. This model enhances energy efficiency and scalability and enables easy expansion with new weights without hardware changes, unlike single-stage models. The first stage operates continuously to detect abnormal signals, activating subsequent stages only when abnormalities are detected. This approach reduces memory consumption by 52%, increases power efficiency, and achieves a modest average AUC improvement of 0.4% over traditional single-stage models.