Implementation of Self-Organized Operational Neural Networks for R-Peak Detection in Holter ECGs

Despite the large number of R-peak detectors that have been proposed in the literature, their robustness and performance levels may significantly deteriorate when low-quality and noisy data are collected from mobile electrocardiogram (ECG) sensors, such as Holter monitors. This problem has recently been tackled by deep 1-D convolutional neural networks (CNNs), which have recently achieved state-of-the-art performance levels in Holter monitors. However, because to their extreme complexity, real-time CNN processing calls for specialized hardware architecture that is parallelized. Instead, they perform worse when they adopt a compact network configuration. This is the expected outcome as earlier studies have demonstrated that the strictly uniform architecture of CNNs—which employs a single linear neuron model—limits their ability to learn. This problem has been tackled by operational neural networks (ONNs), which encapsulate neurons with a range of nonlinear operators in a heterogeneous network topology. In this study, we suggest using generative neurons in 1-D Self-Organized ONNs (Self-ONNs) to enhance peak detection performance while maintaining a sophisticated computational efficiency. The ability of 1-D Self-ONNs to self-organize eliminates the need to identify the optimal operator set for every neuron since every generative neuron may generate the optimal operator during training, which is the main advantage over ONNs. The experimental results show that the proposed 1-D Self-ONNs can achieve better performance than the state-of-the-art deep CNNs with reduced computational complexity, based on the MIT BIH Arrhythmia dataset dataset with over a million ECG beats.

Keywords: Convolutional neural networks (CNNs), Holter Monitors, Operational Neural Networks (ONNs), R-peak detection, MIT BIH Arrhythmia dataset.