Authors:
Leandro Duque Mussio
and
Maria Claudia F. Castro
Affiliation:
Department of Electrical Engineering, Centro Universitário FEI, Brazil
Keyword(s):
PPG, Photoplethysmography, Signal Quality, Artifact, STFT, Deep Learning, CNN, Public Dataset, Small Sample Size, Biomedical Signal Analysis.
Abstract:
Photoplethysmography (PPG) signal analysis has the potential for various medical applications, such as heart rate monitoring, blood pressure estimation, and emerging techniques like diagnosing diabetes and glucose level estimation. However, noise and artifacts, especially motion artifacts, can degrade the quality of PPG signals, making it difficult to extract meaningful features. This research addresses this challenge by investigating the quality of photoplethysmography (PPG) signals using the Short-Time Fourier Transform (STFT) and a deep learning model. The objective is to classify PPG signals as good or bad to eliminate bad signals and increase the accuracy of subsequently derived features. The signals were pre-processed using the publicly available BUT PPG database, consisting of a limited number of smartphone PPG recordings with a low sampling rate (30 Hz), generating spectrographic images used in training a Convolutional Neural Network (CNN) to classify the quality of the signa
ls. Nested cross-validation with five external folds and two internal stratified folds was applied to optimize hyperparameters and assess the model’s performance. The results show the effectiveness of the proposed approach, improving the extraction of features from PPG signals by collecting 94.29% (± 7.82%) of good signals and filtering 80% (± 12.78%) of bad signals.
(More)