Treffer: Cardiac arrhythmias detection framework based on higher-order spectral distribution with deep learning.

• To introduce a pre-processor using a cascaded Variable step size NLMS and Sparse Low-Rank filter (CVSS-NLMS-LRF) method to remove noise from ECG signals. • To perform Higher-order spectral energy distribution image (HSDI) method using two-dimensional Fourier transform of the third order cumulant function. • To classify cardiac arrhythmia detection using the multi-stage dual-swin transformer method with attention learning (MS-DSwin-AL). • To propose integrated average subtraction and standard deviation-based optimizer (IASSD) Method for hyperparameter optimization. • To compare the performances of the proposed model with the existing methods in terms of different performance metrics to prove the performance superiority of the proposed method. In this article, a new framework for arrhythmia identification using higher-order spectral distributions and deep learning approaches has been proposed. The input signal is first pre-processed using the Sparse Low-Rank filter (CVSS-NLMS-LRF) and Cascaded Variable step size-Normalized least mean square algorithm (CVSS-NLMS-LRF) techniques. This method eliminates various types of noise signals from the Electrocardiogram (ECG) signals, such as power line noise, baseline wander noise (BW), and high-frequency muscle artefacts. After pre-processing the signal, the features are selected using a Higher-order spectral energy distribution image (HSDI) obtained using a two-dimensional Fourier transform from the third-order cumulant process. Finally, the multi-stage dual swin transformer with attention learning (MS-DSwin-AL) is proposed to classify cardiac arrhythmias in the input ECG spectral. The Dual Swin Transformer, the Channel and Element-wise Attention Mechanism (CEAM) and the Transitional Module (TM) form the framework of the proposed classification. Additionally, the classification parameters are fine-tuned using an Integrated Average Subtraction and Standard Deviation based Optimizer (IASSD) algorithm. As a result, the proposed model is executed in the Python platform using the MIT-BIH dataset, and the performance is considered in terms of various evaluation metrics. Furthermore, the performance of a proposed model is compared with existing classifiers. The proposed model achieves an accuracy (96.01%), specificity (94%), recall (93.02%), and F1-score (89.14%) higher than the existing classifiers for cardiac arrhythmia classification. As a result, it can be said that the proposed model has a strong chance of identifying cardiac arrhythmias from the provided data. [ABSTRACT FROM AUTHOR]