Photoplethysmographic Signal Quality Assessment: a Comparative Study of Unsupervised and Supervised Neural Networks

We present algorithms to discriminate good quality PPG signals, i.e., free of artefacts and with suitable morphologies, which is fundamental for a correct medical diagnosis. We have investigated two different approaches of unsupervised and supervised learning. The first method is a Self Organizing Map (SOM), trained on entropic and morphological features extracted by BUT PPG signal windows. We then add three new features related to signal quality, namely the Kurtosis Index, Skewness Index and Signal to Noise Ratio (SNR), that we have shown to improve the performances. For the second approach, we have implemented a Multi Layer Perceptron (MLP) neural network. We have compared the results obtained with those in the literature showing superior performance, especially with the MLP approach, achieving an $F_{1} -$score of 96.15%.