HCloud, a Healthcare-Oriented Cloud System with Improved Efficiency in Biomedical Data Processing - Cloud Computing with e-Science Applications - page 177

Filtering Baseline
Wander

Detecting the
QRS Wave

Denoising Others

FIGURE 8.8
Four phases of ECG data processing.

8.3.1.1 ECG Data Features

ECG data is collected from the human body with the frequency range
between 0.05 and 100 Hz, and its amplitude is only several millivolts. Hence,
disturbances from the environment always must be a concern, to avoid
baseline wander of signals, and then QRS wave detection will be performed
after signal denoising. The ECG data process can be divided into four phases:
filtering baseline wander, denoising, detecting the QRS wave, and other
postprocess phases, as shown in Figure 8.8.
For the long-term (24-hour) ECG data, data can be up to 12 MB at a sampling
frequency of 150 Hz. An integer coefficient digital filter transfer formula
specific to the ECG signals is as follows.

2

1
30

1
1

z
z

− 90

() −× −
−

− 87

Hz =z (8.1)

− 3

Then, the filtering baseline wander iteration function will be derived from
Equation (8.1):

() = () −− ( ) + ( )

yn xn xn 90 yn − 3

1 1

() = () − ( ) + yyn ( )

yn yn yn − 90 − 3 (8.2)

2 1 1 2

()

yn

() =− ( ) −

2

yn xn 87

900

where x(n) is the original input signal, and y(n) is the output signal (i.e., filtering
baseline wander data). After testing the execution speed of the serial pro-
gram with the raw data, the results showed that filtering the baseline wander
phase took about 70% of the entire ECG processing time in our existing
HCloud platform [ 20 ], which was the bottleneck of our ECG analysis algo-
rithm, as Table 8.2 shows. If a single machine or serial programming is merely
adopted, the user experience would be worse. Therefore, it is considered to
be the first part of computing in parallel on the Map-Reduce framework.

8.3.1.2 Parallel Programming of the ECG Data Process

According to the analysis presented, the computing overload of the ECG
data process is mainly at the phase of filtering the baseline wander. So,
filtering the baseline wander should be parallelized first. Raw data from the