Recently,it is reported that the cardiovascular diseases have been one of the top killers of human beings.However,the death rate can be largely suppressed if the blood pressure,which reveals the heath condition of cardiovascular system,can be continuously monitored.Moreover,because of the strong fitting ability,the machine learning algorithms have been increasingly applied in blood pressure(BP)prediction researches.Thus,we have developed a learning-based cuffless blood pressure measurement model in this thesis.Specifically,the system takes input of photoplethysmography(PPG)signal from the finger and then extracts the corresponding features,which are further used by the least square support vector machine(LSSVM)models to predict the BP value.Motivated by saving the power consumptions and hardware resources,we have made several improvements in the algorithms,such as mathematical morphological filter,feature extraction,feature selection,feature scaling,outlier removal and support vectors sparseness,which were further verified by the PPG and BP signals from University of Queensland Vital Signs Dataset.Finally,an offline cuffless blood pressure measurement model has been trained.In TI's TMDX5535 eZdsp ultra-low power develop board,we have developed a two-features based cuffless blood pressure measurement prototype.The system consists of PPG sensor,DSP and GUI in PC,of which we mapped all proposed sampling,filtering,feature extractor and LSSVM inference function modules in the DSP platform.The prototype can be used for further development of the wearable BP measurement devices.Based on the measurements from 5 volunteers,the system reports a tolerable error(MAE±STD)of 8.42±10.30 mmHg for SBP and 6.03±7.64 mmHg for DBP. |