Abstract
This study aimed to develop a real-time, noninvasive hyperkalemia monitoring system for dialysis patients with chronic kidney disease. Hyperkalemia, common in dialysis patients, can lead to life-threatening arrhythmias or sudden death if untreated. Therefore, real-time monitoring of hyperkalemia in this population is crucial. We propose a wearable single-lead ECG monitoring system, offering enhanced comfort and feasibility for extended use. The key innovation of this system is the design of a compact, multi-channel convolutional neural network. This model offers high stability, strong performance, and exceptional computational efficiency, making it ideal for seamless integration into wearable devices for real-time monitoring applications. The model automatically extracts features from ECG signals at different frequencies through multiple convolutional channels, eliminating the need for manual feature extraction before data input. Data is input using a non-overlapping sliding window approach, reducing preprocessing complexity while maintaining model performance. We investigated the optimal window length and the number of convolution channels for ECG signal input. Experimental results indicate that the model achieves optimal performance with a 1200 ms window length and four parallel convolutional branches, yielding an accuracy of 98.25% (4.52%), F1-score of 98.31% (3.26%), sensitivity of 98.63% (2.41%), and specificity of 97.88% (5.13%). This system holds significant potential for improving patient monitoring comfort and real-time responsiveness.