Artifact Noise Removal Techniques and Automatic Annotation on Seismocardiogram Using Two Tri-axial Accelerometers

Abstract

Heart disease are ones of the most death causes in the world. Many studies investigated in evaluating the heart performance in order to detect cardiac diseases in the early stage. The aim of this study is to monitor the heart activities in long-term on active people to reduce the risk of heart disease. Specifically, this study investigates the motion noise removal techniques using two-accelerometer sensor system and various positions of the sensors on gentle movement and walking of subjects. The study also ends up with algorithms to detect cardiac phases and events on Seismocardiogram (SCG) based on acceleration sensors. A Wi-Fi based data acquisition system and a framework on Matlab are developed to collect and process data while the subjects are in motion. The tests include eight volunteers who have no record of heart disease. The walking and running data on the subjects are analyzed to find the minimal-noise bandwidth of the SCG signal. This bandwidth is used to design bandpass filters in the motion noise removal techniques and peak signal detection. There are three main techniques of combining data of the two sensors to mitigate the motion artifact: analog processing, digital processing and fusion processing. The analog processing comprises analog ADDER/SUBTRACTOR and bandpass filter to remove the motion before entering the data acquisition system. The digital processing processes all the data using combinations of total acceleration and z-axis only acceleration. The fusion processing automatically controls the amplification gain of the SUBTRACTOR to improve signal quality as long as a signal saturation is detected. The three techniques are tested on three placements of sensors including horizontal, vertical, and diagonal on gentle motion and walking. In general, the total acceleration and z-axis acceleration are best techniques to deal with gentle motion on all placements which improve average systolic signal-noise-ratio (SNR) around 2 times and average diastolic SNR around 3 times comparing to only one accelerometer. With walking motion, overall the ADDER and zaxis acceleration are best techniques on all placements of the sensors on the body which enhance about 7 times of average systolic SNR and about 11 times of average diastolic SNR comparing to only one accelerometer. The combination of two sensors also increases the average number of recognizable systole and diastole on walking corresponding to 71.3 % and 43.8 % comparing toiii only one sensor. Among the sensor placements, the performance of horizontal placement of the sensors is outstanding comparing with other positions on all motions. There are two detection stages to detect events in the SCG for automatic annotation. First, two algorithms including moving average threshold and interpolation are applied to locate the systolic and diastolic phases. Then, based on those identified phases, cardiac events are found in the searched intervals using two outstanding characteristics of the SCG. The two algorithms of phase detection are examined on the stationary data sets of digital processing and horizontal placement. The total acceleration of only one sensor is also calculated for comparison. With moving average threshold algorithm, the average error and missing rates of total acceleration and z-axis acceleration are 1.8 % and 2.1 % respectively which are lower than using one accelerometer (3.6 %). With interpolation algorithm, the average error and missing rates of total acceleration and z-axis acceleration are in the order of 2.3 % and 2.4 % which are still lower than one accelerometer. The average calculation time of the moving average algorithm is lower than the interpolation counterpart. The real-time mode of detection algorithms is also demonstrated on Matlab framework to prove the possibility of practical applications.