| dc.description.abstract | Blood pressure is one of the most important metrics for cardiovascular disease diagnosis and treatment. The conventional measurement using cuff is not convenient and comfortable. In addition, the so-called white coat effect may influence the measurement result. Recently, the wearable devices that use non-invasive approach-es to obtain long-term monitoring of health conditions attract much attention. The wearable devices also helps to realize diagnosis and treatment of telemedicine. In this thesis, we use Electrocardiography (ECG) signal and Photoplethysmography (PPG) signal from a finger device to estimate blood pressure. In the preprocessing phase, the 60 Hz interference and baseline drift are removed from the signals, and bad-quality waveforms are eliminated by a pre-screening function. Then, the feature extraction function applies which extracts the blood pressure related features such as R, S, T peaks of ECG and the extrema of PPG as well as of the first and second deriva-tive PPG signals. Our feature extraction method of ECG R peak is tested for over 100,000 beats in MIT-BIH database, and it finally shows 99.46% accuracy. The weighted waveform decomposition technique is utilized to decompose the PPG pulse into five Gaussian component waves. Their amplitudes, widths and peak positions are also employed as our features. In the data processing phase, we screen out outliers of waveforms and features, calculate the median of feature values. The random over-sampling strategy is adopted to balance the training data set in different regions. Fi-nally, four machine learning methods are used and compared. The feature expansion scheme also shows performance improvement. Multi-layer perceptron (MLP) and convolutional neural network (CNN) are well integrated to enhance the regression model of BP. Our testing set contains 85 subjects and 255 records with distribution satisfying the AAMI requirements. The estimation performance achieves root mean square error of 10.23 (mmHg) in systolic blood pressure and 8.22 (mmHg) in diastolic blood pressure without subject-based correction. Compared to the conventional work in a similar testing condition, better estimation results are achieved. | en_US |