| dc.description.abstract | The detection of beat positions and types within ECG signals stands as a fundamental, noninvasive,
and economically viable approach to preliminarily evaluate heart rate and cardiac well-being in
patients. This necessitates precise beat detection, particularly over short time frames. Nevertheless,
the reliability of automatic beat detection is hindered by diverse noise sources and complex signal
characteristics, while traditional algorithms have demonstrated certain shortcomings. This thesis
aims to address the challenges of determining ECG beat positions by deep learning technology
and accurately estimating heart rate values. There are three main steps in the training phase of the
thesis which include: data preprocessing, training deep learning model, and model evaluation. In
the preprocessing process, the signal must pass through signal resample, band-pass filter, and
baseline removal. Then the signal was transformed and annotations were relabeled. The Unet1D
model will be utilized for training beat detection. After training, the model will be carefully
evaluated and be embedded into the tflite format for running in Raspberry. Finally, a web server
connects the user information will be built for further evaluation. The training result was 95%
accurate for AHA and 85% for NSTDB datasets. The time for predicting one four-second segment
is around 0.8 seconds. Experimental results show the viability of the proposed method for
automated beat detection and heart rate estimation. Furthermore, the practical implementation
highlights that the proposed method's utility extends beyond ECG data analysis, showcasing its
potential for actual device creation. | en_US |
