A cascaded approach of group sparse mode decomposition and deep neural network for heart rate estimation using reference signal-less PPG signal

•The proposed GSMD-based method effectively removes MA from a PPG and doesn’t require a reference signal.•Proposed a CNN-Bi-LSTM-Attention model for robust HR estimation from PPG signals.•Applied Bayesian optimization for enhanced accuracy and computational efficiency.•Achieved MAE of 2.42 BPM, vali...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Measurement : journal of the International Measurement Confederation Jg. 246; S. 116546
Hauptverfasser: Pankaj, Maan, Pratibha, Kumar, Manjeet, Kumar, Ashish, Komaragiri, Rama
Format: Journal Article
Sprache:Englisch
Veröffentlicht: Elsevier Ltd 31.03.2025
Schlagworte:
Bayesian optimization
Bi-LSTM
CNN
Deep learning
Group sparse mode decomposition
Heart rate estimation
K-fold cross-validation
PPG signal
Deep learning
Bi-LSTM
CNN
PPG signal
Bayesian optimization
K-fold cross-validation
Group sparse mode decomposition
Heart rate estimation
ISSN:0263-2241
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:•The proposed GSMD-based method effectively removes MA from a PPG and doesn’t require a reference signal.•Proposed a CNN-Bi-LSTM-Attention model for robust HR estimation from PPG signals.•Applied Bayesian optimization for enhanced accuracy and computational efficiency.•Achieved MAE of 2.42 BPM, validated on six diverse datasets and 90 participants.•Framework enables real-time HR monitoring with reduced computational complexity. Subjects with cardiovascular disease (CVD) require continuous heart rate (HR) monitoring and must take timely treatment to prevent emergencies. The use of wearable wristband-type smart healthcare devices for routine check-ups of the cardiac state is becoming common. Among the other smart health monitoring wearable devices, the framework uses a Photoplethysmogram (PPG) optical sensor to monitor HR in real-time. However, these devices are vulnerable to the presence of motion artifacts of the user. Earlier techniques used a PPG signal and accelerometer signal to remove the motion artifacts for accurate and reliable estimation of HR from the raw PPG signal. Techniques that employ a single PPG sensor are preferred to reduce hardware complexity and power consumption. The general framework for most techniques consists of data acquisition, pre-processing of data, motion artifacts removal stage, HR estimation stage, and post-processing stage. The performance of the traditional signal processing technique relies on the post-processing stage. Machine learning and deep network-based algorithms overcome the need for a post-processing stage for accurate estimation of HR. Different pre-processing and motion artifact stages of obtaining a clean PPG signal to train a model increase the computational burden and become a deterrent to use in wearable devices. The main objectives of the proposed work are to improve the accuracy of HR estimation and to reduce the need for reference signals. This work proposes a group-sparse mode decomposition (GSMD) based pre-processing stage and a 1-D deep neural network-based HR estimation stage to estimate HR accurately in real-time. The proposed framework integrates the convolutional neural network, bidirectional long short-term memory (CNN-Bi-LSTM)-Attention mechanism to extract valuable features from the clean PPG signal obtained using GSMD. The variation in the HR during different exercises is predicted using a deep neural network (DNN) that consists of a CNN-Bi-LSTM and attention layer. In addition, the performance of the proposed model is improved using Bayesian optimization. The results obtained under the proposed framework yielded a mean absolute error (MAE) of 2.42 BPM. The proposed GSMD and optimized deep neural network-based framework can effectively predict HR in real-time. This work optimized the proposed 1-D model parameter using a Bayesian optimization algorithm and a 10-fold cross-validation approach to find the generalized accuracy of the trained model. The proposed system provides an affordable, efficient, and fully connected monitoring solution for cardiac patients. The proposed framework, evaluated with six publicly available datasets, comprises 90 subjects and 76,735 eight-second windows recorded with different demographic protocols.
ISSN:0263-2241
DOI:10.1016/j.measurement.2024.116546