On a Feature Extraction and Classification Study for PPG Signal Analysis

Photoplethysmography (PPG) is a low cost, non-invasive optical technology to detect the volumetric changes of blood circulation at the surface of skin. While the medical indication of components of PPG signals in the form of pulse wave are not yet fully understood, it is vastly agreed that they carry valuable pathophysiological information related to the cardiovascular system. Going beyond just dealing with frequency and time domain features of the pulse wave, as well as the first and second derivatives of the wave commonly seen in many of the relevant work, we applied a K-MEANS improved algorithm for feature extraction based on selected time domain parameters: K1 (systolic area), K2 (diastolic area) and K (entire pulse wave area). The extracted characteristic waveforms under the same light intensity could achieve average confidence level of 90% or higher. The stationary wavelet transform was adopted to further analyze the characteristic waveform by calculating the wavelet entropy; We then trained a Probability Neural Network (PNN) model using the wavelet entropy and other time domain characteristic parameters. It is found that the trained PNN model performs well in analyzing characteristic waveform to distinguish between health condition and severe arterial stenosis.