A Mechanism for Data Quality
Estimation of On-Body Cardiac
Sensor Networks
Sunghoon Ivan Lee*
Charles Ling*
Ani Nahapetian*†
Majid Sarrafzadeh*
*Computer Science, UCLA
†Computer Science, CSUN
Copyright: UCLA Wireless Health Institute
Wireless Health Institute (WHI) - UCLA
• Campus Community
–
–
–
–
–
–
–
School of Medicine
Medical Center
School of Engineering
School of Nursing
School of Public Health
College of Letters & Science
Anderson School of Management
• Unique approach
– End-to-end integration from
sensing to medical informatics to
call center
– Develop and verify new
healthcare methods and services
– Establish standards for efficacy,
reliability, interoperability, and
security
Cardiac-Monitoring Sensors
• BANs or PANs involve various types of wearable and noninvasive sensors on body
– Cardiac Sensor, EMG, EEG, Glucose sensor, Motion sensor, etc.
• Cardiac monitoring sensors are the most common and widely
used sensors
– For example, in the survey on wearable sensor-based systems for
health monitoring [17], 90% of the introduced systems involve cardiac
monitoring sensors.
Copyright: UCLA Wireless Health Institute
3
Problems with Wearable Cardiac
Sensors
• Sensors often suffer from high level of noise
• Types of possible noise include
i. Channel noise produced by human body [11, 19]
ii. Noise caused by environments [5]
iii. Noise from loose physical contact of the sensor
node to the human body
• Motion artifacts have the most significant
degradation effect on the quality of sensor data,
especially when the subject is highly mobile
(e.g., at-home remote health care applications).
4
Motivation Example
• This signal was obtained from one of our participants wearing
an on-body ECG sensor (Alive Heart Monitor [1]) at wrist.
Copyright: UCLA Wireless Health Institute
5
Motivation Example
• Therefore, multiple sensors are often
mounted at different parts of the body to
provide higher data quality in a highly mobile
environment.
Copyright: UCLA Wireless Health Institute
6
Data Quality Information
• For continuous and pervasive health monitoring
environment, data quality information can be
very useful
– Automatically detecting sensors producing clean data
• In the field of medicine, extra manpower is required to
manually filter out polluted data [20]
– Improving continuous monitoring of patients
– The data quality information can be also used to
optimize the resources the pervasive system.
• This is a very important issue for pervasive systems
Copyright: UCLA Wireless Health Institute
7
Objective
• Our work introduces a mechanism for data
quality estimation of a BAN composed of
cardiac sensors specifically considering
resource scarceness
Copyright: UCLA Wireless Health Institute
8
Considered BAN Structure
Sensor node #1
Aggregator
Sensor node #2
• The mechanism
considers a BAN
structure that all cardiac
sensors transmit data to
a single aggregator
Body Sensor Network
Copyright: UCLA Wireless Health Institute
9
Summary of the Proposed Mechanism
• STEP 1: Local Data Quality Estimation
Sensor node #1
Aggregator
Sensor node #2
• Individual sensors filter out most of
the normal events and recognize any
abnormal events (motion artifact noise
+ health hazardous events)
Body Sensor Network
Copyright: UCLA Wireless Health Institute
10
Summary of the Proposed Mechanism
• STEP 2: Global Data Quality
Estimation
Sensor node #1
Aggregator
Sensor node #2
• Aggregates information about the local
data quality and fuses the information
to estimate the data quality of the
overall BAN.
Body Sensor Network
Copyright: UCLA Wireless Health Institute
11
Local Data Quality Estimation
• Detect any abnormal events in data generated
from a single cardiac sensor
• This method is based on a well known fact
that amplitude of cardiac signal and interpulse interval (IPI) variability are effect
bedside measurements to detect any
abnormal events [4], [8], [9].
Copyright: UCLA Wireless Health Institute
12
The overview of the local data quality
estimation
• Digital filters include [16]
i.
ii.
iii.
•
An integer coefficient band-pass filter
A derivative filter combined with an amplitude square process
Moving window integrator
Peak detection logic locates cardiac cycles [18]
Copyright: UCLA Wireless Health Institute
13
The overview of the local data quality
estimation
• Using the location of cardiac cycles, we extract
the time length of each cardiac cycle (i.e., IPI)
• We define this high pass filtered IPI time series as
IPI variation, and denote it as v[n]
Copyright: UCLA Wireless Health Institute
14
The overview of the local data quality
estimation
• The time series of average amplitude of
cardiac cycle is denoted as a[n ]
Copyright: UCLA Wireless Health Institute
15
Irregular Fluctuation
• Now we have new time series v[n ] and a[n]
• The proposed mechanism focuses on discarding
most of normal cardiac cycles based on observing
irregular fluctuation
– Irregular fluctuation describes erratic movements in a
time series that follow no recognizable or regular
patter [5].
– Health hazardous events carry irregular fluctuation in
v[n ] and/or a[n ]. [14]
– Our observation verifies that motion artifact noise
also carries irregular fluctuation in v[n] and/or a[n]
Copyright: UCLA Wireless Health Institute
16
Pattern of normal cycles
• We define the pattern of normal cycles as the
degree of variation in normal v[n] and a[n] within
a window size of N assuming Gaussian
distribution
– It is actually known that the distribution of normal
cardiac cycles are usually skewed rather than Gaussian
[12].
– However, these models are very complicated, and
usually evolve over time.
– Since our objective is to discard most of normal
cardiac cycle rather than accurately model the
distribution, Gaussian is a good approximation
Copyright: UCLA Wireless Health Institute
17
Pattern Learning Process
• The pattern learning process involves computing the
mean and the std. dev. of normal cycles for a window
size of N.
– Learning process bounded by O(N).
– It requires local memory to store N numbers.
• The proposed mechanism determines that the newest
cardiac cycle of index n has high data quality if
v   v   v  v[n]  v   v   v  & a   a   a  a[n]  a   a   a 
note that δ can be adaptively chosen as a result of the training process
Copyright: UCLA Wireless Health Institute
18
Local Data Quality Estimation
• The output of the local estimation, which is then
transmitted to the aggregator is
1,
Q[n ]  
0,
if
v   v   v  v[n]  v   v   v  & a   a   a  a[n]  a   a   a 
otherwise
• In our experiment, we asked participants to sit on
a chair without any movements to acquire N=20
normal cycles.
• Then, δ is chosen such that all N normal cycles
satisfy the above inequality.
Copyright: UCLA Wireless Health Institute
19
Global Data Quality Estimation
• The data fusion process is performed at the aggregator side.
STEP 1: Temporal synchronization based on the minimum sampling rate
for real-time purpose.
(sensors may have different sampling rate)
Synchronization
t1[m]
t2[m]
t3[m]
Copyright: UCLA Wireless Health Institute
20
Global Data Quality Estimation
• Then, tk[m] is defined as
tk[m] = Qk[n]
for m within each cardiac cycle.
(mathematical details provided in the paper)
• Intuitively, it is a time series for the kth sensor
that shows the quality of a cardiac cycle in a
synchronized sampling rate
Copyright: UCLA Wireless Health Institute
21
Global Data Quality Estimation
• STEP 2: we fuse the synchronized tk[m] to
estimate the data quality t[m] of the overall
BAN using a majority voting.
Copyright: UCLA Wireless Health Institute
22
Experimental & Simulation Results
• Experiment
– Conducted on 4 participants to show that the
proposed mechanism recognizes the noise created
by motion artifacts.
• Simulation
– Simulation on multi-variable cardiac data (from
PhysioNet) to address that the mechanism can
recognize the noise created by health hazardous
events such as hearth arrhythmia.
Copyright: UCLA Wireless Health Institute
23
Experiment
• 4 Participants
• 3 off-the-shelf cardiac sensors (i.e., K = 3)
– CHEST: Alive Heart Monitor
– LEG: Vernier EKG Sensor
– FINGER: Nonin Onyx 9560 SpO2 Sensors
Copyright: UCLA Wireless Health Institute
24
Experiment
• Set of actions that simulates the average daily activity of a person
based on the American Time User Survey (ATUS) [21].
–
–
–
–
–
Walking
Sitting down with no movement
Sitting down while moving upper limbs
Bending down to pick up an object
Standing up while moving upper limbs
• Participants performed the actions for 10 seconds and rest for
another 10 seconds. Then, this combination of action and rest is
repeated 3 times
– clearly distinguishing the noise from normal signal
• We manually annotated all cardiac cycles to be either normal or
abnormal (noise), and compared the detection results (i.e., tk[m])
against this ground truth annotation for each sensor data.
Copyright: UCLA Wireless Health Institute
25
Illustrative Example of Experimental
Results
Copyright: UCLA Wireless Health Institute
26
Experimental Results
rd: the overall detection rate
rfa: false abnormal rate
rfn: false normal rate
• In average, the detection rate for the abnormal cycles can be detected at the rate
of 90.56% when the data was fused. The false abnormal and false normal rates
were 3.89% and 25.11%, respectively.
Copyright: UCLA Wireless Health Institute
27
Simulation
• The database used in this simulation is the MGH/MF
Waveform Database from PhysioNet [10].
– Due to the limitation and safety issues in recruiting
participants with severe cardiac ailments who are likely to
undergo a health hazardous cardiac problem during the
experiment
• Includes
–
–
–
–
Three ECGs
An arterial pressure
A pulmonary arterial pressure
A central venous pressure signal.
Copyright: UCLA Wireless Health Institute
28
Simulation
• Two interesting observations
1. All local signals had the same IPI time series
since none of the sensors is locally distorted due
to motion artifacts.
2. The average ratio of the number of normal to
abnormal cardiac cycles is 99.1%.
• We investigate detection rate for normal and abnormal
cycles separately.
Copyright: UCLA Wireless Health Institute
29
Simulation Results
N’a: # detected abnormal
Na: # actual abnormal
rfn: false normal rate
N’n: # detected normal
N’n: # actual normal
rfa: false abnormal rate
• In average, the detection rate for the abnormal cycles is 100% and the detection
rate for the normal cycles is 87.8%.
Copyright: UCLA Wireless Health Institute
30
Summary
• We introduced an efficient mechanism for estimating data quality
of a BAN composed of cardiac sensors.
– Low complexity
•
The proposed method employs
– Local data quality estimation
– Global data quality estimation
•
We presented experimental results based on three off-the-shelf
cardiac sensor devices in order to detect motion artifact noise.
• We also presented simulation results to detect health hazardous
events using the proposed mechanism.
Copyright: UCLA Wireless Health Institute
31
Thank you
• Questions?
• Please feel free to reach me at silee@cs.ucla.edu
Copyright: UCLA Wireless Health Institute
32