Photoplethysmography Signal Quality Assessment Using Neighbour Edge Restricted Horizontal Visibility Graph and Machine Learning Classifiers

Photoplethysmography (PPG) signals are vital for monitoring pulse rate, blood pressure, and more, but they are prone to motion artefacts and noise, leading to unreliable data. Assessing PPG signal quality is crucial for reliable healthcare and accurate medical diagnoses. By transferring the PPG time domain signal to a Horizontal Visibility Graph (HVG) network and combining extracted features from HVG with machine learning algorithms, we can classify the PPG signal into clean (or high quality) and noisy. We have proposed a new version of HVG called Neighbour Edge Restricted Horizontal Visibility Graph (NERHVG) by invoking some extra conditions for joining edges in HVG for PPG signal quality assessment (SQA). We have used the average degree (AD) of graphs extracted from HVG, and NERHVG algorithms as features in 3 different machine learning classifiers such as Random Forest (RF), Gaussian Naive Bayes (GNB), Decision Tree (DT) to classify 4 standard untrained PPG datasets (DS). The classifier models DT, RF, GNB associated with graph feature AD of HVG, NERHVG algorithms are named as: DT-HVG, DT-NERHVG, RF-HVG, RF-HVG, RF-NERHVG, GNB-HVG and GNB-NERHVG. After all the performance of HVG and NERHVG algorithms using AD feature are compared over the mentioned classifier models. It is observed that the NERHVG algorithm outperformed the HVG algorithm with the AD feature in all 4 datasets with a maximum accuracy of: $99.09\%, 95.03\%, 96.56 \%$ and $84.63\%$ using the GNB classifier.