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

While a number of R-peak detectors have been created, their performance may be significantly impacted when handling noisy, low-quality data from mobile ECG sensors, such as Holter monitors. Even though deep 1-D convolutional neural networks (CNNs) have recently produced state-of-the-art results, their high complexity and need for specialized parallel hardware for real-time processing can limit performance, particularly with compact network configurations. Because CNNs only use one linear neuron model, their learning capacity is limited, leading to this constraint. To tackle this problem, operational neural networks (ONNs) integrate neurons with several types of nonlinear operators in a network architecture that is heterogeneous. The goal of this work is to improve R-peak detection performance in 1-D Self-Organized ONNs (Self-ONNs) while maintaining computing efficiency through the use of generative neurons. Because each generating neuron in a 1-D Self-ONN learns its ideal configuration through adaptation, the Self-Organizing feature eliminates the need for human operator set selection. Our experimental results, utilizing the MIT-BIH Arrhythmia dataset, which contains over a million ECG beats, reveal that 1-D Self-ONNs outperform state-of-the-art deep CNNs in terms of both performance and computational economy.