Light Weight Graph Neural Network Models for PPG Signal Quality Assessment

Photoplethysmography (PPG) signals are widely used to measure vital signs, such as pulse rate, blood oxygen saturation, blood pressure, and respiration rate, in a non-invasive and cost-effective manner. However, motion artifacts often distort PPG signals, compromising their reliability in wearable and ambulatory monitoring contexts. Recently, Graph Neural Networks (GNNs) and visibility graph techniques, which transform time-series signals into complex graph networks, have shown promise for biosignal analysis. This paper proposes two novel Signal Quality Assessment (SQA) models, HVGCN and NERHVGCN, which use the Horizontal Visibility Graph (HVG) and Neighbour Edge Restricted Horizontal Visibility Graph (NERHVG) algorithms, respectively, in combination with Graph Convolutional Network (GCN) architectures. Both models employ node degree and local clustering coefficients in node feature vectors to assess signal quality by classifying PPG signals as noise-free (acceptable) or noisy (unacceptable). Our results demonstrate that the NERHVGCN model achieves an overall accuracy of 97.59% on four unseen PPG datasets, outperforming the HVGCN model, which achieved 96.98%. Both models are lightweight, with sizes of 4.4 kB (NERHVGCN) and 4.3 kB (HVGCN), respectively. These findings highlight the potential of visibility graphs and GNNs for effective SQA applications, providing a robust, resource-efficient solution for wearable health monitoring. To the best of our knowledge, this is the first work utilizing visibility graphs and GNNs for PPG SQA. The complete code is available at: github.com/Zahir-Khan98/GNN-PPG-SQA.