Interpreting ECG Data by integrating statistical and artificial intelligence tools

Abstract

The use of an automated system integrating data conditioning,
statistical methods, and artificial intelligence tools to summarize and
interpret high-frequency physiological data such as the
electrocardiogram is investigated. The development of a methodology and
its associated tools for real-time patient monitoring and diagnosis is
accomplished by using the commercial programming environments MATLAB and
G2, a real-time knowledge-based system (KBS) development shell. Data
interpretation and classification is performed by integrating
statistical classification methods and knowledge-based techniques with a
graphical user interface that provides quick access to the analysis
results as well as the original data. A KBS was developed that
incorporates various statistical methods with a rule-based decision
system to detect abnormal situations, provide preliminary interpretation
and diagnosis, and to report these findings to the healthcare provider

Full text

Interpreting ECG Data by
Integrating Statistical and
Artificial Intelligence Tools
Combining Knowlege-Based Systems with Statistical
Methods for More Robust Intelligent Patient Monitoring
T
he interpretation of physiological data
obtained from patient monitoring de
-
vices has created new challenges for
healthcare professionals in their efforts to
extract useful information from the raw
data. Online monitoring equipment used
in modern medical facilities generate
large amounts of data that must be inter-
preted quickly and accurately. The avail-
ability of portable monitoring devices for
ambulatory home care and paramedic ap-
plications introduce situations in which
an expert may be unavailable to interpret
patient data. Conditioning of data is nec-
essary to remove noise and motion arti-
facts via digital signal processing (DSP)
techniques such as filtering via wavelet
decomposition and reconstruction.
In this article, the use of an automated
system integrating data conditioning, sta
-
tistical methods, and artificial intelli
-
gence tools to summarize and interpret
high-frequency physiological data such
as the electrocardiogram (ECG) is inves
-
tigated. The development of a methodol
-
ogy and its associated tools for real-time
patient monitoring and diagnosis is ac
-
complished by using the commercial pro
-
gramming environments MATLAB and
G2, a real-time knowledge-based system
(KBS) development shell. Data interpre
-
tation and classification is performed by
integrating statistical classification meth
-
ods and knowledge-based techniques
with a graphical user interface (GUI) that
provides quick access to the analysis re
-
sults as well as the original data. A KBS
was developed that incorporates various
statistical methods with a rule-based de
-
cision system to detect abnormal situa
-
tions, provide preliminary interpretation
and diagnosis, and to report these find
-
ings to the healthcare provider.
An Overview of Intelligent
Patient Monitoring
Recent applications of systems for
medical data acquisition and analysis in
-
clude a real-time system for monitoring
and diagnosis of cardiovascular system,
an intensive-care monitoring and decision
support system, a blood-glucose predic-
tion system, a remote patient monitoring
system, and a modular approach to a pa-
tient monitoring system [1-5]. Each of
these applications focuses on a specific
area, such as data transmission, acquisi-
tion, or software architecture. However,
these systems typically focus on one spe-
cialty of patient monitoring. A robust
combination of computer and human di-
agnostic strengths, specifically, quantita
-
tive and qualitative reasoning, is therefore
needed for a comprehensive intelligent
patient monitoring system.
Computer-based patient monitoring
systems should therefore be capable of in
-
tegrating many sources of data both in
real-time and on an “as-available” basis.
The most important patient variables, or
vital signs, are monitored continuously.
These include blood pressure, pulse rate,
respiration rate, and temperature. Auto
-
mated monitoring systems may further as
-
sist healthcare providers by performing
fast, complex statistical analysis of the vi
-
tal signs in real time.
Statistical methods for detecting
changes in physiological parameters are
included in a field generally known as sta
-
tistical process monitoring (SPM) or sta
-
tistical quality control (SQC). The most
widely used and popular SPM techniques
involve univariate methods, which are
proven, simple, and easy to implement.
This simplified approach to patient moni
-
36 IEEE ENGINEERING IN MEDICINE AND BIOLOGY January/February 20020739-5175/02/$17.00©2002IEEE
Eric Tatara and Ali Cinar
Dept. of Chemical and Environmental Engineering,
Illinois Institute of Technology
© 1989-99 TECH POOL STUDIOS, INC.

toring requires a healthcare provider to
continuously monitor perhaps dozens of
different univariate charts, which sub
-
stantially reduces his ability to make ac
-
curate assessments about the state of the
patient. Multivariate statistical process
monitoring (MSPM) methods are typi
-
cally concerned with measuring
directionality of data from a multivariable
space as well as the magnitude and varia
-
tion of changes as opposed to univariate
methods that only monitor the magnitude
and variation of single variables [6-9].
MSPM techniques reduce the amount of
raw data presented to a healthcare pro
-
vider and provide a concise set of statistics
that describe the patient’s health.
As with patient monitoring, recent de
-
velopments in the field of telemedicine in
-
volve the use of computers for data
acquisition, transmission, and processing.
The advent of high-speed digital commu
-
nication lines has further increased both
the amount of data that can be exchanged
as well as the quality of the data. Conse-
quently, current research trends in
telemedicine involve structured architec-
ture for telehealth systems [10], as well as
the development and integration of new
equipment and communication standards
[11-13]. Instead of relying on local equip-
ment and expertise, healthcare providers
can now enlist the help of doctors from
around the world to solve difficult diagno-
sis problems. Physiologic signals may be
transmitted in real-time to an expert for
evaluation or stored for later analysis.
Graphical information such as radiologi
-
cal images and video feeds of the patient
may also be transmitted from remote sites.
A healthcare provider can view the patient
on a monitor, review physiologic
real-time data as well as patient records,
and recommend specific treatments, ef
-
fectively providing a level of care that was
previously unavailable. However, the task
of interpreting transmitted patient data
still remains a formidable challenge.
Healthcare personnel providing assis
-
tance via a telemedicine system may be
aided by more advanced data analysis
techniques that are not currently being
used in patient monitoring. More robust
analysis tools for patient monitoring will
include aspects of MSPM for data pro
-
cessing as well as intelligent tools for han
-
dling of qualitative patient data.
Knowledge-based systems (KBS) or
expert systems are computer systems de
-
signed in an attempt to emulate the deci
-
sion-making capabilities and knowledge
of a human expert in a specific field. KBSs
consist of a knowledge base, decision
rules, and an inferencing engine. The
knowledge base is comprised of a set of
data or facts pertaining to a specific pro
-
cess. Decision or production rules are typi
-
cally of the IF/THEN variety that operate
on data acquired from the patient and the
knowledge base. The inferencing engine is
responsible for making conclusions about
information from the database, patient
data, and production rules. KBSs are useful
for solving problems that can only be done
by human experts or are repetitive in na
-
ture. Applications in the medical field in
-
clude diagnosis assistance, intelligent
patient monitoring, and alarm handling.
Knowledge-based systems are devel
-
oped using languages capable of symbolic
processing such as LISP, PROLOG, or
more specialized KBS development soft
-
ware such as G2 by Gensym Corporation,
Cambridge, MA. The advantage of these
software over software commonly used in
engineering such as FORTRAN is the
ease of programming knowledge qualita-
tively. Additionally, code written by one
programmer may be easily understood by
another, without significant knowledge of
the language. While there are a large num-
ber of available software packages for ad-
vanced numerical analysis, the great
majority of these software are not written
in a way that is compatible with software
for KBSs.
A robust system for patient monitoring
must utilize the powerful MSPM tech
-
niques that have been developed for analy
-
sis of patient data as well as features of
KBS for qualitative reasoning. Such a hy
-
brid system will use the best of both tech
-
niques. A large portion of the available
commercial software for MSPM and KBS
are used off-line; that is, data are processed
and analyzed after an event has occurred or
the process has ceased. A real-time patient
monitoring and diagnosis system is neces
-
sary to detect and diagnose serious prob
-
lems as they occur, in order to take
immediate corrective measures.
Signal Conditioning
and Feature Detection
The most fundamental types of patient
data that are available include physical
parameters that may be measured directly.
Computerized data acquisition systems
are capable of displaying and storing the
measure variable as well as real-time
analysis. Signals that are clear and have a
high signal-to-noise ratio are relatively
easy to interpret, both by a computer and a
human healthcare provider. Often, how
-
ever, physiological signals are corrupted
by large amounts of noise that make inter
-
pretation difficult. Noise contamination
of physiological signals is typically in the
form of motion artifacts caused when the
patient moves. Sensors that measure
physical movement such as respiration
are especially susceptible to motion arti
-
facts. Other types of noise include electri
-
cal interference or powerline noise as well
as baseline drift, which are usually the re
-
sult of poorly designed or faulty instru
-
mentation.
One of the most well studied and under
-
stood physiological signals is the measured
electrical potential related to the beating of
the heart. The ECG relates the observed
ionic current on the skin to events that oc
-
cur in heart. The ECG is typically mea
-
sured by placing various electrodes on the
chest as well as the arms and/or legs.
Arrhythmias, or abnormal heart rhythm,
can be detected by examining the ECG
[14-16]. Data are obtained from the
MIT/BIH arrhythmia database, which is
considered the most popular data source
for evaluation of ECG analysis tools.
By analyzing the structure of the ECG
obtained from a patient, a healthcare pro-
vider may determine the root cause of the
disease. Each segment of the ECG signal
represents distinct physiological func-
tions and can therefore be used as a diag-
nostic aid. Comparison of new ECG
waveforms from a patient to a reference
standard is very beneficial in making a di
-
agnosis. While the detection of each of the
waves of the ECG signal is important for
diagnosis of cardiological pathologies,
the QRS segment is by far the most impor
-
tant in monitoring the heartbeat.
Before detailed analysis of physiologi
-
cal signals is possible, signal conditioning
is often required to remove sources of
noise. Noise is often manifested as
high-frequency components in the signal,
while the lower-frequency components
contain more pertinent information and
are thus of greater interest for analysis. It
is possible to remove the high-frequency
components of a signal using Fourier
analysis, in which the signal is divided
into several sinusoids of different fre
-
quency. By summing the low-frequency
components and excluding the high-fre
-
quency components, the noise can be ef
-
fectively removed.
Since the Fourier transform converts
time information to frequency informa
-
January/February 2002 IEEE ENGINEERING IN MEDICINE AND BIOLOGY 37

tion, the time at which certain frequencies
manifest themselves in the signal is lost.
The wavelet transform allows the analysis
of the frequency components of a signal
with respect to their temporal relationship
to one another [17, 18]. Wavelets are
functions of finite duration whose mean
value is equal to zero over the duration of
the function. Wavelet techniques for fil
-
tering are similar to those of the Fourier
transform in removing the high-frequency
noise components of the signal. A func
-
tion is fitted to a signal by scaling and
stretching the original, or mother wavelet,
to match the waveform of the signal.
Typically, a wavelet is chosen that some
-
what resembles the morphological fea
-
tures of the signal to be filtered. The
continuous wavelet transform (CWT)
yields a set of wavelet coefficients that,
multiplied by the scaled and time-shifted
mother wavelet, produces a set of wave
-
lets whose sum is the original signal. The
advantage of using wavelets for filtering
is that the scaled and time-shifted daugh
-
ter wavelets are a better representation of
local phenomena in the original signal,
whereas Fourier analysis is more suitable
for periodic signals.
Most physiologic signals are
nonstationary, meaning that they do not
follow a statistical distribution around a
constant mean value over time. Therefore,
it is necessary to obtain derived character
-
istics or features of physiologic signals for
use with statistical monitoring techniques.
Knowledge of the location of characteristic
features of physiological signals such as
the ECG are clearly important in diagnosis
of pathologies. While feature detection and
analysis may be directly implemented in a
patient-monitoring KBS, the program
-
ming expense would be enormous. The
benefits of using a KBS are in the areas of
diagnosis support and should use data that
is derived from low-level signals, not the
signals themselves.
Typical features of interest in physio
-
logical signals are peaks and troughs in
the waveform and the temporal distance
between the different waveform features.
Detection of the QRS complex in the ECG
waveform is one of the most common and
beneficial types of feature detection. Each
QRS represents a heartbeat and is used in
determining the R-R interval, the distance
between two consecutive R-peaks in the
ECG waveform. If the R-R interval is
known, the heart rate (HR) may be easily
determined. By observing the heart rate
variability (HRV) over time, healthcare
providers can make diagnosis based on
the behavior of the heart and the effect on
the other physiological systems.
Monitoring
The Shewhart control chart, developed
in the 1920s by Walter A. Shewhart, is a
graphical means of displaying quality
characteristics calculated from industrial
process variables [6]. Quality characteris
-
tics or variables are statistics that are de
-
rived from measured medical data
obtained from patient monitoring devices.
The purpose of the Shewhart chart is to
show a plot of the desired quality variable
of interest versus time as well as a region
bounded by upper and lower control lim
-
its that define the in-control process (Fig.
1). The area between the lower and upper
control limits represents the natural varia
-
tion of the data. The Shewhart charts help
to distinguish between significant varia
-
tions of normally distributed variables
caused by abnormal conditions and varia
-
tions occurring naturally in the process. If
each observation of the measured quality
characteristic falls between the control
limits on the control chart, the data are in a
region of normal conditions, or consid
-
ered to be in statistical control. If an obser
-
vation falls outside of the limits, the pro
-
cess is out of control. By using the
Shewhart control charts to monitor patient
variables, one may assign a cause to ab
-
normal conditions by observing which
variable or variables have exceeded the
control limits.
The model of the control chart is de
-
scribed by the center line (CL), upper con
-
trol limit (UCL), and lower control limit
(LCL). The center line is defined as the
mean of the variable of interest. The UCL
and LCL are determined by the overall
variability, or variance, of the in-control
data. The standard deviation σ reflects the
scatter of data about the mean, and there
-
fore indicates a nominal range for the data:
UCL x S=+3
LCL x S=−3
where
x
is an estimate of the mean µ and S
is the unbiased estimate of the actual stan
-
dard deviation σ. Limits of plus or minus
three times the standard deviation provide
confidence limits of over 99.7 percent.
This confidence interval indicates that
any observation that falls within the
3-sigma limits is very likely to be in con-
trol and any observation that falls outside
the limits is very likely to be out of con-
trol.
Shewhart charts only provide informa-
tion for monitoring the magnitude and
variation of single variables. However,
the in-control region of many physiologi
-
cal systems is defined by several vari
-
ables, thereby rendering the univariate
control chart inappropriate for monitor
-
ing, when used exclusively [6-9].
Univariate charts as shown in Fig. 1 repre
-
sent quality characteristics in only
one-dimensional space, in terms of the
possible value of the characteristic. To
create a better representation of the state
of the patient, a multidimensional space is
used. Figure 2 shows data from the
univariate control charts such as Fig.1 in
the form of a bivariate control chart with
each axis representing a process variable.
The in-control variability of the process
can now be described by a control ellipse
whose axes are proportional to the
in-control standard deviation of the re
-
spective variables as shown in Fig. 2. Any
observation that falls within the ellipse is
considered in control and any observation
outside of the ellipse is out of control. Be
-
cause the ellipse extends beyond the UCL
and LCL of each of the variables some
-
38 IEEE ENGINEERING IN MEDICINE AND BIOLOGY January/February 2002
LCL Mean UCL
Sample
Number
Sample Number
UCL
Mean
LCL
X
1
X
2
2. Bivariate control chart.
R-R Interval
1
0.9
0.8
0.7
0.6
0.5
UCL
LCL
0 50 100 150 200 250 300
Time
1. Univariate Shewhart chart.

what on one axis and lies significantly
well within the UCL and LCL on another
axis, there will be a disagreement between
the multivariate and univariate methods
for determining when the data are
in-control. The area within the ellipse is
distributed as
χ
α ,2
2
where α represents the
confidence interval used [8]. This control
chart is referred to as the chi-square con
-
trol chart. This technique may be applied
to higher-order systems with more than
two variables as well by plotting the
Hotelling’s
T
2
statistic described below.
The control space becomes an ellipsoid in
three dimensions and a hyper-ellipsoid in
larger dimensions.
The multivariate
χ
2
chart improves the
validity of the space of in-control opera
-
tion defined by the univariate charts, al
-
though with a sacrifice. When plotting
one process variable versus another, tem
-
poral information is lost. An out of control
observation may be obtained, but unless
the time at which it occurred and the tem
-
poral relation to the other observations is
known, the information is not of much
use. Additionally, while one may observe
the data in a two- or three-dimensional
space, making inferences about the state
of a process from multidimensional data
becomes incredibly difficult. The
χ
2
sta-
tistic is a multivariate statistic that mea-
sures the statistical distance of an
observation from its in-control mean:
χµ µ
21
=− −
−
() ()xx
T
Σ
The
[]1× p
observation vector x contains a
single observation of all measured p vari
-
ables while µ is a similarly structured vec
-
tor of the corresponding means for each
process variable. Σ is the
[]pp×
in-control
covariance matrix that is calculated from
a set of n observations, whose elements
are the variances between each variable,
and
()⋅
T
denotes the transpose operation.
The
χ
2
statistic may then be calculated for
each observation of p variables and plot
-
ted similarly to the univariate control
chart. The LCL is zero and the UCL is
χ
α , p
2
. Observations that fall above the
UCL indicate an out-of-control process.
Through the use of the
χ
2
control chart,
one must assume that the means and vari
-
ances for the required variables are
known. As with the univariate control
chart, estimates of the quality variables
are usually required. A new quality statis
-
tic called the Hotelling’s
T
2
statistic is
used when the population means and
covariance are unknown:
T
T21
=− −
−
() ()xx S xx
where x once again represents a single
[]p ×1
observation vector of p variables.
The vector
x
represents the estimated
mean for each variable, and S is the esti
-
mated covariance matrix. The
T
2
statistic
may be calculated and plotted for each
new observation (Fig. 3). The UCL for the
T
2
statistic is:
UCL
pn
nn p
F
T
pn p
2
2
1
=
−
−
−
()
()
,,α
where
F
pn pα ,,−
is the F-distribution pa
-
rameter with a confidence of
()1−α
per
-
cent and
(, )pn p−
degrees of freedom.
Statistical Classification
A new set of derived variables can be
computed from the ECG waveform in or
-
der to better detect ventricular fibrillation
and ventricular tachycardia [19]. The per
-
cent of time above and below thresholds
(PTABT) and the thresholds crossing in
-
tervals (TCIs) are two such variables that
can be obtained in real-time from an ECG
signal. PTABT represents the time that
beat peaks are located outside statistically
determined threshold limits. TCI is the in-
terval between points that are outside of
the threshold limits. The standard devia-
tion of TCI (SDTCI) represents the regu-
larity of the signal. Arrhythmias with
irregular patterns such as ventricular fi-
brillation can be more easily detected us-
ing SDTCI because the signal is very
irregular, inflating the SDTCI statistic.
Figure 4 shows a plot of SDTCI versus
PTABT with several groups of
arrhythmias plotted. Sinus (normal)
rhythm is distinctly separate from ventric
-
ular fibrillation data. This plot is of little
value used alone, since a healthcare pro
-
vider must once again rely on knowledge
of where normal data is separated from
abnormal data. However, used with more
powerful statistical classification meth
-
ods, the PTABT and SDTCI statistics are
of more practical use.
Classification of physiological signals
using MSPM techniques is possible using
data obtained directly from the patient in
real time. The original work by Tian and
Tompkins [19] proposed arbitrary
univariate limits for classification of heart
arrhythmias from PTABT and SDTCI
data. If a point falls within the proposed
limits, it is given the respective classifica
-
tion. However, univariate limits are not an
accurate representation of the true classifi
-
cation boundaries. By using a control el
-
lipse based on the
χ
2
distribution, the re
-
gion associated with both the sinus and
ventricular fibrillation clusters becomes
better defined as shown in Fig. 5. The clas
-
sification boundaries may be determined a
priori from a database of patient histories,
or modified in real time based on newly ob
-
tained data. Classification is then per
-
formed by examining the statistical
distance of the new observation from the
bounded regions. Observations that fall
within the classification boundary may be
confidently classified accordingly. Those
observations that fall outside of any of the
known regions may be labeled as unclassi
-
fied or classified based on the closest dis
-
tance to a known category.
A discriminant function
$
y
may be used
to determine the statistical classification
of an observation into one of two normally
distributed populations
π
1
and
π
2
with
similar covariance structure [7]:
[]
$$
y
T
T
==−
−
ax x x S x
012
1
0p
where
x
0
is a
[]21×
array containing the
two observed variables ,
x
1
and
x
2
are
[]21×
arrays containing the means for
January/February 2002 IEEE ENGINEERING IN MEDICINE AND BIOLOGY 39
0 100 200 300 400 500 600 700 800
99% UCL
50
45
40
35
30
25
20
15
10
5
0
T
2
Time
T
2
Chart
3. T
2
control chart
Sinus
VFib 1
VFib 2
50
45
40
35
30
25
20
15
10
5
0
SDTCI
0
5101520253035
PTABT
4. SDTCI versus PTABT for sinus and
ventricular tachycardia rhythms.

both variables in each population, and
S
p
is the
[]22×
pooled covariance matrix be
-
tween two populations with
n
1
and
n
2
ob
-
servations:
SS
S
p
n
nn
n
nn
=
−
−+ −




+
−
−+ −




1
12
1
2
12
2
1
11
1
11
()()
()()
which is an unbiased estimate of the actual
covariance matrix Σ.
S
1
and
S
2
are the
sample covariance matrices from the re
-
spective populations. The populations re
-
fer to bivariate or multivariate clusters of
data that are obtained from any physiolog
-
ical parameters. The discriminant func
-
tion is compared to a midpoint between
the two populations given as
$
()
m
yy
=
+
12
2
where
y
T
11
=
$
ax
y
T
22
=
$
ax
.
If
$$
ym≥
, the observation is classified
as belonging to
π
1
.If
$$
ym≤
, the observa
-
tion is classified as belonging to
π
2
.By
using this discriminant function, a com
-
puter algorithm can be easily imple
-
mented to classify data obtained in real
time since the computational requirement
is small. The statistical distance from each
state can be monitored over time using the
discriminant function to identify trends in
the data [9]. While the discriminant func
-
tion may not classify all points that do not
fall within the nominal operating region, a
trend chart will alert healthcare providers
of an impending transition into an abnor
-
mal state of health. Statistical classifica
-
tion may be extended to more than two
populations and to populations that are
non-normally distributed and/or have dif
-
ferent covariance structures. Extension to
higher-order dimensions of data greatly
enhances the diagnostic capabilities of a
healthcare provider, as it is very difficult
to mentally visualize the data in more than
two dimensions.
Patient Monitoring KBS
G2 is a graphical knowledge base de
-
velopment environment for creating intel
-
ligent real-time applications. Typical
engineering applications include moni
-
toring and diagnosis, alarm handling,
scheduling, design, and simulation. Ap
-
plications developed in G2 are called
knowledge bases (KBs) and may be saved
in a format that allows periodic mainte
-
nance or modifications. A KB contains
workspaces that resemble windows avail
-
able in most graphical operating systems,
upon which objects are organized.
Workspaces may be arranged in a hierar
-
chical structure with one workspace at the
top of the hierarchy and multiple
workspace below, thus providing means
for organization of data. Workspaces con
-
tain all of the rules, variables, and objects
that are used when developing a KB.
KB structure may be further organized
into several different KBs called modules.
G2 modules provide a means for separat
-
ing different kinds of objects when devel-
oping a large KB as well as ease in
reusability of code. By utilizing the modu-
lar structure, information is better orga-
nized providing ease of code maintenance
and KB operation. Modules such as those
containing definitions of patients may also
be shared with other KB systems, eliminat-
ing the need to rewrite large portions of
code. Graphical representation of data in
G2 is performed through several different
types of charts that are customizable by the
user. The displays may be placed on any
workspace and the shape, position, and
colors of the display may be modified as
needed by the user. Charts have real-time
capability and can display any type of nu
-
merical or symbolic data. Controls such as
buttons and slider bars update the values of
the user-specified variables that pertain to
individual controls.
The KBS performs data collection and
processing, graphical display of data via
user-specified formats, and diagnosis as
-
sistance via heuristics and rules. The envi
-
sioned application area is in care wards
where one healthcare provider is respon
-
sible for monitoring of several patients si
-
multaneously, although the techniques
are readily transferable to other medical
monitoring situations. The top-level inter
-
face menu displays information for each
of the patients that are being monitored as
shown in Fig. 6. The top-level screen is an
overview to the entire ward and alerts the
healthcare provider when an abnormal sit
-
uation arises. Detailed information can be
40 IEEE ENGINEERING IN MEDICINE AND BIOLOGY January/February 2002
Sinus
99% Limit
95% Limit
VFib 1
99% Limit
95% Limit
VFib 2
99% Limit
95% Limit
70
60
50
40
30
20
10
0
–10
SDTCI
0 5 10 15 20 25 30 35 40
PTABT
5. Bivariate control chart for SDTCI and PTABT statistics.
6. Top level KBS interface.
7. KBS detail-level screens showing pa
-
tient information, vital signs, and
charting.

immediately accessed through the top-
level screen by choosing one of the avail
-
able selections for each patient.
The healthcare provider can inspect
more detailed information on any of the
lower level screens. Several detail screens
are shown in Fig. 7. Personal information
contains items such as the patient’s age,
sex, and supervising physician. Buttons
are available for navigation through other
screens containing patient medical his
-
tory, medications, lab results, etc.
Charting for various physiological vari
-
ables can be viewed by selecting the de
-
sired variable. An ECG waveform for the
selected patient is shown in Fig. 7. Addi
-
tionally, indicators for each of the moni
-
tored vital signs can be viewed
simultaneously. The variables shown in
Fig. 7 include pulse rate, respiration, tem
-
perature, blood pressure, blood oxygen
saturation, and cardiac output. Each ob
-
servation of the current variable is shown
by an indicator bar with a background of
several different colors. The indicator dis
-
plays a digital display of variable as well
as a moving graphical indicator The color
regions behind the indicator provide
healthcare providers with a quick over-
view of the state of the patient. Colors
were chosen to be easily understandable:
green represents the nominal in-control
region, yellow for the region bounded by
the 95% and 99% confidence intervals,
and red for the region outside of the 99%
confidence limit. By glancing at the vari
-
able of interest, a healthcare provider can
quickly determine if the variable is within
the normal operating condition. Most
alarm conditions may be easily detected
by someone observing the patient moni
-
tor, yet this is a difficult task for the moni
-
toring of only one patient. The KBS
contains a rule base that monitors the indi
-
cator variables and sounds an alarm when
any of the multivariate statistics exceed
their limits. The rule base is capable of ex
-
amining several variables simulta
-
neously, thereby providing diagnosis
based on multiple observed symptoms.
When an alarm is sounded, the rule base
provides a list of potential causes for the
alarm as well as a list of corrective mea
-
sures to be taken.
Conclusions
A framework for an intelligent patient
monitoring and diagnosis assistance sys
-
tem was developed by investigating sig
-
nal processing and statistical techniques
and their integration with KBSs. By inte
-
grating statistical and knowledge-based
techniques, a system can be developed
which is more robust that a system con
-
sisting of only the individual techniques.
The KBS reduces the amount of informa
-
tion that healthcare providers must pro
-
cess in time critical situations and permits
them to give more attention to treating the
patient. The graphical interface of the
KBS provides easy access to patient infor
-
mation and monitoring data as well as di
-
agnosis support to healthcare providers
who are not specialists in treating certain
diseases.
Eric Tatara received his B.S. and M.S. in
chemical engineering from Illinois Insti
-
tute of Technology in 1997 and 1999, re
-
spectively. He is currently a Ph.D.
candidate in the Process Modeling, Moni
-
toring and Control Group in the
Department of Chemical Engineering at
Illinois Institute of Technology. His re-
search interests include statistical process
control, artificial intelligence, and moni-
toring applications in biomedical engi-
neering.
Ali Cinar received his Ph.D. in chemical
engineering from Texas A&M University
in 1976. He is currently professor of
chemical engineering and Dean of the
Graduate College at Illinois Institute of
Technology. His research interests in
-
clude statistical process monitoring and
fault diagnosis; supervisory control of
process operations by integrating system
theoretical, statistical, and artificial intel
-
ligence based techniques; and modeling
and monitoring applications in biomedi
-
cal engineering.
Address for Correspondence: Eric
Tatara, Illinois Institute of Technology,
Department of Chemical and Environ
-
mental Engineering, 10 West 33rd Street,
Chicago, IL 60616. Tel: +1 312 567 3522.
Fax: +1 312 567 8874. E-mail:
tataeri@iit.edu.
References
[1] A-R. Tanha and D. Popovic, “An intelligent
monitoring system for patient health care,” in
Proc. IEEE EMBS 19th Annu. Int. Conf. vol. 3,
Chicago, IL, 1997, pp. 1086-1087.
[2] V. Moret-Bonillo, M.J. Cabrero-Canosa, and
E.M. Hernandez-Pereira, “Integration of data, in
-
formation and knowledge in intelligent patient
monitoring,” Expert Syst. Appl., vol. 15, pp.
155-163, 1998.
[3] J.J. Liszka-Hackzell, “Prediction of blood glu
-
cose levels in diabetic patients using a hybrid AI
technique,” Comput. Biomed. Res., vol. 32, pp.
132-144, 1999.
[4] S.H. Park, J.H. Park, S.H. Ryu, T. Jeong, H.H.
Lee, and C.H. Yim, “Real-time monitoring of pa
-
tients on remote sites,” in Proc. IEEE EMBS 20th
Annu. Int. Conf., Hong Kong, 1998, pp. 1321-1325.
[5] T. Sukuvaara, M. Sydanmaa, H. Nieminen, A.
Heikela, and E. Koski, “Object-oriented imple
-
mentation of an architecture for patient monitor
-
ing,” IEEE Eng. Med. Biol. Mag., vol. 12, pp.
69-81, Dec. 1993.
[6] D.C. Montgomery, Introduction to Statistical
Quality Control. New York: Wiley, 1997.
[7] R. Johnson and D. Wichern, Applied
Multivariate Statistical Analysis. Upper Saddle
River, NJ: Prentice-Hall, 1998.
[8] T. Kourti and J.F. MacGregor, “Multivariate
SPC methods for process and product monitoring,”
J. Qual. Technol., vol. 28, pp. 409-428, 1996.
[9] A. Raich and, A. Cinar, “Statistical process
monitoring and disturbance diagnosis in
multivariable continuous processes,” AIChE J.,
vol. 42, pp. 995-1009, 1996.
[10] S. Warren, R.L. Craft, R.C. Parks, L.K.
Gallagher, R.J. Garcia, and D.R. Funkhouser, “A
proposed information architecture for telehealth
system interoperability,” in Proc. IEEE EMBS
21st Annu. Int. Conf., Atlanta, GA, 1999, p. 702.
[11] K. Shimizu, “Telemedicine by mobile com-
munication,” IEEE Eng. Med. Biol. Mag., vol. 18,
pp. 32-44, 1999.
[12] J. Bai, Y. Zhang, D. Shen, L. Wen, C. Ding,
Z. Cui, F. Tian, B. Yu, B. Dai, and J. Zhang, “A
portable ECG and blood pressure telemonitoring
system,” IEEE Eng. Med. Biol. Mag., vol. 18, pp.
63-70, 1999.
[13] J. Presedo, S. Barro, D. Castro, S.
Fernandez-Delgada, S. Fraga, M. Lama, and J.
Vila, “Intelligent telemonitoring of critical care
patients,” IEEE Eng. Med. Biol. Mag., vol. 18, pp.
80-88, 1999.
[14] J.D. Bronzino, The Biomedical Engineering
Handbook. Hartford, CT: CRC, 1995.
[15] D. Dubin, Rapid Interpretation of ECGs.
Tampa, FL: Cover Publishing, 1996.
[16] G. Naccarelli, Cardiac Arrhythmias: A Prac
-
tical Approach. Mount Kisco, NY: Futura, 1991.
[17] G. Strang and T. Nguyen, Wavelets and Fil
-
ter Banks. Wellesley MA: Wellesley-Cambridge,
1996.
[18] Y.T. Chan, Wavelet Basics. Boston, MA:
Kluwer, 1995.
[19] L. Tian and W.J. Tompkins, “Time domain
based algorithm for detection of ventricular fibril
-
lation,” in Proc. IEEE EMBS 19th Annu. Int.
Conf., vol. 3. Chicago, IL, 1997, pp. 374-377.
January/February 2002 IEEE ENGINEERING IN MEDICINE AND BIOLOGY 41