Enhanced ECG Signal Detection using ID-CNN with Attention Mechanism

Abstract

Hyperkalemia, a potentially life-threatening condition marked by elevated potassium levels in the blood, is often detectable through abnormalities in electrocardiogram (ECG) signals. Despite this, detecting hyperkalemia remains challenging due to the limited sensitivity of existing models in identifying subtle ECG changes associated with varying potassium levels. To address this challenge, we propose an enhanced ECG signal detection model employing a 1D Convolutional Neural Network (CNN) integrated with an attention mechanism. This model architecture features three 1D CNN layers for deep feature extraction, followed by an attention mechanism that dynamically adjusts the weight of these features to improve detection sensitivity. Utilizing the MIMIC-IV dataset, which encompasses diverse ECG recordings, our model demonstrated superior performance in detecting hyperkalemia compared to traditional CNN models. By focusing on the most relevant aspects of ECG signals, our approach significantly enhances detection accuracy and sensitivity. Additionally, comparative analysis of lead performance revealed that even with fewer leads, such as a 2-lead ECG, the model achieved comparable accuracy, making it suitable for deployment in wearable devices.