Frequency and Uncertainty Driven Deep Learning Approach to Segment Electrocardiogram Signals for Effective Heart Parameters Estimation

Accurate classification of electrocardiogram signals is reliant on accurate heart rhythm parameters detection, which requires effective electrocardiogram segmentation at the beat level. In this study, we propose and evaluate integrating temporal, frequency, and uncertainty-informed deep learning approaches for classifying electrocardiogram signals into three main cat-egories: PQ, QRST, and TP. Utilizing PhysioNet's QT public database, we preprocess the data, including noise filtering, gap removal, and normalization, to prepare it for deep learning model input. We employ multilayer deep learning architectures, integrate them with additional features such as short-time fourier transform and approximate entropy to determine uncertainty-enhanced classification performance. Through extensive experimentation and five-fold cross-validation, we analyze the impact of layer duplication and additional features on classification accu-racy. Our results demonstrate that our proposed methodologies achieve up to 96 % accuracy in classifying Electrocardiogram (ECG) signals, providing valuable insights for improving heart rhythm parameter detection and suggesting avenues for further research in ECG signal processing.