Expanding PPGFeat for Multi-Dataset PPG Analysis,Comprehensive Biomarker Extraction and Machine learningfor Cardiovascular Risk Prediction

Sammanfattning

The number one cause of mortality worldwide is cardiovascular disease and early detection of cardiovascular risk factors can greatly reduce the risk of severe complications. Photoplethysmography (PPG) is a method to non-invasively measure changes in blood volume that can readily be implemented into wearable medical devices. This is beneficial as PPG can be used to estimate numerous cardiovascular parameters, however the underlying physiological mechanisms behind it are still not fully understood. PPGFeat is a MATLAB toolbox designed to support researchers in analysing raw PPG data by providing signal preprocessing and fiducial point extraction with a graphical user interface. This thesis presents the development of an updated version of PPGFeat that expands on the old toolbox by adding support for more diverse datasets, automatic signal quality assessment and expanding the number of PPG derived biomarkers. The methods for loading data were successfully extended to accommodate a larger number of datasets without negatively affecting the fiducial point extraction. Furthermore, the increased number of biomarkers were used to train two classifiers to determine hypertension status with over 55% accuracy across four classes. Finally, while automatic signal quality assessment was implemented in the toolbox, the resulting index is only marginally more accurate over using the mean signal quality of the dataset which limits its potential use.