Abstract

The PPG signal presents considerable promise as a non-invasive technique across various applications. However, effectively utilizing this signal in real-world scenarios demands meticulous handling to identify and rectify disturbances within the photo-plethysmography (PPG) signal. Among the methodologies explored, integrating time-frequency spectra with a hybrid deep learning model, such as convolutional – long short term memory neural network model (CNN-LSTM), has emerged as a promising approach. Yet, prevalent methods often rely on Fourier-based algorithms for extracting time-frequency spectra, which are prone to energy leakage issues. To surmount this limitation, decomposition methods like Variational Mode Decomposition (VMD) coupled with the Hilbert transform offer a compelling solution. In this study, we propose a novel algorithm leveraging VMD and Hilbert transform to extract time-frequency spectra as features for a convolutional neural network model (CNN). Unlike studies employing Fourier-based time-frequency spectra and the hybrid CNN-LSTM model, this approach adopts a simpler architecture, relying solely on a CNN model. This simplicity owes to the efficacy of VMD and Hilbert transform in feature extraction, streamlining the computational process without sacrificing accuracy. Remarkably, our method yields high-performance outcomes, achieving accuracy, precision, and recall of 0.91, 0.95, 0.88, respectively on the MIMICIII dataset. These results underscore the robustness and effectiveness of our proposed methodology, offering promising avenues for enhanced utilization of the PPG signal in diverse biomedical applications. By amalgamating advanced signal processing techniques with deep learning models, our approach contributes to the advancement of non-invasive biomedical signal processing, potentially healthcare monitoring and diagnosis.