ANNet: A Lightweight Neural Network for ECG Error Detection in Edge Sensors

Abstract

With the usage of Internet of Things (IoT) edge sensors, we’re going to implement a less massive neural network to detect Electro-Cardiogram (ECG) anomaly. The network comprises of both Long Short Term Memory cells (LSTM) as well as Multi-Layer Perceptron in aggregation. The MLP layer receives the characteristics produced from instants of heart rate and the LSTM is fed with a series of coefficients of the denoised signal which are denoised using moving average filter constitutes the characteristics of ECG beat. By simultaneously training the blocks, the entire network is driven to learn unique characteristics that complement one another for decision-making. The network's accuracy, computational complexity were evaluated using data from the MIT-BIH arrhythmia database. To address the dataset's class imbalance, we increased the dataset using the SMOTE network training technique. The network's categorization accuracy averaged 98% over several database records. The suggested solution outperforms existing approaches in terms of computational complexity, and it has the advantage of standalone operation in the edge node without requiring constant wireless communication, which makes it perfect for Internet of Things wearable devices.