Spectral feature extraction using the UFIR iterative smoother algorithm for ECG signal classification

Abstract

Heart diseases are the leading cause of death worldwide. One effective non-invasive method for diagnosing heart-related conditions is the analysis of Electrocardiogram (ECG) recordings. These recordings capture the shape of the primary ECG waves, facilitating the automated detection of various pathologies. However, the accuracy of these measurements can be affected by noise or artifacts that occur during the ECG acquisition process. Although many techniques have been proposed to address this issue, there remains a need to improve the precision of automatic ECG signal detection and classification. Our study aimed to reduce noise and extract features from ECG signals associated with arrhythmia, congestive heart failure, and normal sinus rhythm. We evaluated the performance of the Unbiased Finite Impulse Response (UFIR) smoother by comparing it with other techniques, using the root Mean Square Error (RMSE) under various noise levels. Our findings highlighted the significant advantages of the UFIR technique. In addition, we conducted tests using Analysis of Variance (ANOVA) and Kruskal–Wallis analysis to explore the time–frequency domain features obtained from the Short-Time Fourier Transform (STFT) of the states produced by the UFIR smoother. These enhancements improved the performance of the classification model. The results indicated that machine learning techniques based on optimized neural networks achieved impressive metrics, including an F1-score of 0.96, a precision of 95.65%, an accuracy of 93.39%, and Cohen’s kappa of 0.90. Notably, the states of ECG signals provided by the UFIR smoother offer features that could significantly enhance the diagnosis of these pathologies.