IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 44, NO. 9, SEPTEMBER 1997 891
A Patient-Adaptable ECG Beat Classifier
Using a Mixture of Experts Approach
Yu Hen Hu,* Senior Member, IEEE, Surekha Palreddy, and Willis J. Tompkins, Fellow, IEEE
Abstract—We present a “mixture-of-experts” (MOE) approach
to develop customized electrocardigram (ECG) beat classifier in
an effort to further improve the performance of ECG processing
and to offer individualized health care. A small customized
classifier is developed based on brief, patient-specific ECG data.
It is then combined with a global classifier, which is tuned to a
large ECG database of many patients, to form a MOE classifier
structure. Tested with MIT/BIH arrhythmia database, we observe
significant performance enhancement using this approach.
Index Terms— ECG beat classification, MIT/BIH database,
mixture of experts, neural network, patient adaptation.
I. INTRODUCTION
C
OMPUTERIZED electrocardiography is now a well-
established practice, after several years of significant
progress. Many algorithms have been proposed over years for
electrocardiogram (ECG) beat detection and classification. In
a clinical setting, such as an intensive care unit, it is essential
for automated systems to accurately detect and classify elec-
trocardiographic signals on a real-time basis. Since several
arrhythmia are potentially dangerous and life threatening, if
not detected within a few seconds to a few minutes of its
onset, automated electrocardiographic monitoring assumes a
challenging role. Several algorithms have been proposed in
the literature for detection and classification of ECG beats and
reported results, that leave room for improvement. They in-
clude signal processing techniques; such as frequency analysis,
template matching, and other parameter extraction methods.
Artificial neural networks were also employed to exploit their
natural ability in pattern-recognition tasks for successful clas-
sification of ECG beat [2], [3], [6]–[8], [23]–[25], [28]–[31].
One major problem faced by today’s automatic ECG anal-
ysis machine is the wild variations in the morphologies of
ECG waveforms of different patients and patient groups.
An ECG beat classifier which performs well for a given
training database often fails miserably when presented with
a different patient’s ECG waveform. Such an inconsistency
in performance is a major hurdle preventing highly reliable,
fully automated ECG processing systems to be widely used
clinically.
Manuscript received September 13, 1995; revised May 5, 1997. Asterisk
indicates corresponding author.
*Y. H. Hu is with the Department of Electrical and Computer
Engineering, University of Wisconsin, Madison, WI 53706 USA (e-mail:
hu@engr.wisc.edu).
S. Palreddy and W. J. Tompkins are with the Department of Electrical and
Computer Engineering, University of Wisconsin, Madison, WI 53706 USA.
Publisher Item Identifier S 0018-9294(97)06116-8.
One obvious approach to alleviate this problem is to use as
much training data as possible to develop the ECG classifier.
This is the approach taken by all the vendors of ECG pro-
cessing devices: A large in-house ECG database is developed
and maintained to test each ECG processing algorithm to be
incorporated into the product. However, such an approach
suffers several pitfalls.
1) No matter how large this database may be, it is not
possible to cover every ECG waveform of all potential
patients. Hence, its performance is inherently limited.
2) The complexity of the classifier grows as the size of the
training database grows. When a classifier is designed
to correctly classify ECG from millions of patients
(if it ever becomes possible), it has to take numerous
exceptions into account. The result is a complicated
classifier which is costly to develop, maintain, and
update.
3) It is practically impossible to make the classifier learn to
correct errors during normal clinical use. Thus, it may be
rendered useless if it fails to recognize a specific type of
ECG beats which occurs frequently in certain patient’s
ECG records.
The answer, we believe, is to allow the classifier to be
“patient-adaptable.” That is, to let the classification algorithm
adaptable to the special characteristics of each patient’s ECG
records. For example, we may include the training algorithm
and the database used to develop the classifier to be delivered
to the users, so that the classification algorithm can be fine-
tuned to each patient. Unfortunately, this is impractical for
several reasons.
• While it is possible to turn over training algorithms and
databases to the users in an academic environment, it
is unlikely that any commercial ECG machine vendor
is willing to risk revealing their proprietary information
to their competitors. Moreover, in-house database often
contains millions of ECG records which could be costly
to distribute.
• Users often do not want to be bothered by implementation
details of an ECG algorithm. Thus, few users will be able
to take advantage of this patient-adaptation feature even
if it is available.
• Even if a user is willing to perform the patient cus-
tomization, he or she still have to provide sufficient
number of patient-specific training data in order to per-
form patient-adaptation. Manually editing ECG record is
a time consuming, labor intensive task. Hence, the size of
0018–9294/97$10.00 1997 IEEE
892 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 44, NO. 9, SEPTEMBER 1997
patient-specific training data must be tightly controlled.
In this study, we propose a novel approach to patient-
adaptation while avoiding these difficulties: 1) We do not
require the factory-trained ECG classifier to provide training
algorithms or training databases. Instead, all we need is that
this classifier gives both its classification results, as well as
an estimate of posterior probability of the feature vector as is
drawn from each particular class. Hence, no company propri-
etary information is needed. 2) A patient-specific classifier will
be developed using an automated procedure, without human
supervision. 3) Only a brief manually edited patient ECG
record (2–5 min) is needed to achieve significant performance
improvement.
This proposed approach is based on three popular artificial
neural network (ANN)-related algorithms, namely, the self-
organizing maps (SOM), learning vector quantization (LVQ)
algorithms, along with the mixture-of-experts (MOE) method.
SOM and LVQ together are used to train the patient-specific
classifier, and MOE is a paradigm which facilitates the com-
bination of the two classifiers (original and patient-specific)
to realize patient-adaptation. In MOE, the two classifiers are
modeled as two experts on ECG beat classification. The
original classifier, called the Global expert (GE) in this work,
knows how to classify ECG beats for many other patients
whose ECG records are part of the in-house, large ECG
database. The patient-specific classifier, called the local expert
(LE) in this work, is trained specifically with the ECG record
of the patient. A gating function, based on the feature vector
presented, dynamically weights the classification results of the
GE’s and the LE’s to reach a combined decision. The process
is analogous to two human experts arriving at a consensus
based on their own expertise.
Section II reports the results of literature survey and
Section III discusses data acquisition with preprocessing.
Section IV discusses the proposed algorithms and the
development of experts. Section V reports the results of the
classifier on the database records and discusses the results.
Section VI is a summary of the findings of this paper.
II. P
RELIMINARIES
A. ECG Beat Classification Techniques
Automated ECG beat classification was traditionally per-
formed using a decision-tree-like approach, based on various
features extracted from an ECG beat [1], [4], [5], [13], [20],
[22]. The features used include the width and height of QRS
complex, RR interval, QRS complex area, etc. One of the
difficulties is that these features are susceptible to variations of
ECG beat morphology and temporal characteristics. As such,
the classification rate reported in these earlier efforts are rather
moderate.
Artificial neural networks (ANN’s) have been widely ac-
cepted for pattern recognition tasks. Their abilities to learn
from examples and extract the statistical properties of the
examples presented during the training sessions, make them
an ideal choice for an automated process that imitates human
logic. Several efforts have been made to apply ANN’s for
the purpose of ECG beat detection and classification. Previ-
ous reported efforts include [2], [3], [6]–[8], [23]–[25], and
[28]–[31].
Hu et al. [7] reported the development of an adaptive
multilayer perceptron (MLP) for classification of ECG beats.
They have achieved an average recognition accuracy of 90%
in classifying the beats into two groups; normal and abnormal.
In an attempt to classify the beats into 13 groups according to
the MIT Database annotations, they have reported an average
recognition accuracy rate of 65%. An hierarchical system of
the MLP networks which first classify the beat into normal
or abnormal, and then classify it into the specific beat type, is
developed, which improved the recognition accuracy to 84.5%.
B. Self-Organization Map (SOM) and Learning
Vector Quantization (LVQ)
SOM and LVQ are both clustering based algorithms pro-
posed by Kohonen [14], [15]. SOM is an unsupervised on-line
clustering technique. In SOM, each cluster center (prototype
or code word) is represented by the weights of a neuron which
is assigned to a coordinate in the feature map. The SOM
training algorithm forces adjacent neurons in the feature map
to respond to similar feature (input) vectors. In a way, this
feature map is analogous to the spatial organization of sensory
processing areas in the brain. Let
be denoted as the
weights (code word) or the
th neuron in SOM during the time
instant
, the weights of SOM then are updated according to
the following simple formula:
(1)
is the so-called neighborhood kernel, which determine
the size of neighborhood of the
th neuron within which all
neighboring neurons will be updated in response to the present
feature vector
. Initially, the neighborhood is large. The
size reduces as clustering converges, until no neighboring
neurons will get updated.
LVQ is a supervised, clustering-based classification tech-
nique which classifies a feature vector
according to the
label of the cluster prototype (code word) into which
is
clustered. Classification error occurs when the feature vectors
within the same cluster (hence, assigned to the same class
label) are actually drawn from different classes. To minimize
classification error, the LVQ algorithm fine tunes the clustering
boundary between clusters of different class labels by modi-
fying the position of the clustering center (prototype or code
word). This method is called “learning vector quantization”
because this clustering based classification method is similar to
the “vector quantization” method used for signal compression
in the areas of communication and signal processing.
According to Kohonen, there are three different LVQ algo-
rithms, called LVQ1, LVQ2, and LVQ3 developed at subse-
quent stages to handle classification problems with different
natures. In this study, the optimized learning-rate LVQ1 and
LVQ3 algorithms were used for the training and fine-tuning of
the code book respectively. In LVQ1, for a given input vector
, a code word is found such that
(2)
HU et al.: PATIENT-ADAPTABLE ECG BEAT CLASSIFIER 893
The code word is then updated as follows:
(3)
where
if the classification is correct [i.e., and
have the same class label] and , otherwise.
is a time-varying learning rate. Other code words in the code
book remain unchanged. LVQ3 differs from LVQ1 in how
the code words are updated: Assuming that
falls within
a window between two adjacent clusters with corresponding
code words
and . Suppose that and belong to the
same class, and
and belong to different classes, then both
these code words will be updated in LVQ3:
(4a)
(4b)
On the other hand, if both
and belong to the same class
as
, and fall in a window centered at the cluster
boundary of these two classes, then
(5)
The optimal value of
depends on the size of the window,
being smaller for narrower windows. This algorithm is self-
stabilizing, and optimal placement of the
does not change
in continual training.
Software packages of both SOM and LVQ are available
in the public domain,
1
and the application of these packages
to the ECG beat classification problem is straight forward.
The adaptation parameters in these packages (SOM
PAK and
LVQ
PAK) were carefully fine tuned while developing the
classifiers. As such, the development of the code book and
eventually decision boundary can be made completely trans-
parent to the user. Moreover, performance obtained using these
package is very competitive compared to other approaches. In
this research work, we first apply SOM to a set of training
feature vectors. The resulting code book (prototypes) then will
be submitted to the LVQ
PAK to facilitate fine tuning and
classification.
III. M
IXTURE OF EXPERTS (MOE)
This user adaptation problem bears certain resemblance to
the incremental learning problem in that new data are to
be incorporated to improve existing classifier’s performance.
However, the black-box model of the existing classifier pre-
vents us from directly modifying the classifier structure as
incremental learning algorithms do. Instead, we propose a
different method called the MOE, to circumvent this problem.
The MOE approach was proposed by Jacobs et al. [9]–[12],
[16], [26], [27]. The basic notion is that linear combinations
of several statistical estimates can perform better than any
individual estimate. This strategy is not new. It is a well
known fact that a panel of experts often arrive at a better
diagnosis than any single expert, because each expert is able
to contribute from his/her own expertise.
1
University of Helsinki, Finland, URL: ftp://cochlea.hut.fi/pub/
The basic idea is to leave the existing black-box classifier
intact. Instead, we use the given small, user-specific training
data set to develop a LE classifier. Then we invoke a modified
MOE approach to combine these two classifiers, hoping to
achieve better performance.
To apply the MOE approach to solve the customization
problem, we employ two experts: a GE and a LE. The GE
represents the ECG beat classifier developed in factory. Thus,
it is trained to classify all types of ECG beats present in the
in-house ECG database. The LE represents a specialized ECG
beat classifier, trained on a small segment of annotated ECG
beats taken from the specific patient. As such, the GE and the
LE are endowed with complementary knowledge bases, and
can work together to reach a better decision than any one can
reach individually.
The expert network is a combination of the GE and LE
classifiers. Let
and be the output (row) vectors of
the two respective GE and LE classifiers. Each element of each
vector indicates the degree of proximity of an unknown ECG
beat to a predefined ECG beat class (category). In the MOE
method, the combined
th output vector of both the experts
is given by
(6)
where
is the input feature vector, , are the
weighting vectors for each expert from a gating network and
are defined by
(7)
where
’s are the weight vectors of the gating network. Note
that
.
Theorem 1: Define
, and
, , to be the subregion in the feature space where
the classifier
makes correct classification of and let
be defined the same way. Assume and
, then
(8)
Proof: We need only to prove that if both
and
misclassify a given feature vector , then cannot
give correct classification on
. Since the correct classifica-
tion output
, the combined output , and individual
classifier output
and are all binary vectors of the
same dimension, if both classifiers misclassify a given feature
vector
which belongs to class , we must have, for the th
elements of these binary vectors
where “ ” is the “exclusive-OR” operator in Boolean algebra.
Since from (7),
, we conclude ,if
, and ,if .
Hence,
. In other words, must also
misclassify the same feature vector
regardless the choice
894 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 44, NO. 9, SEPTEMBER 1997
of and . This is to say, if , and if
, then .
The implication of Theorem 1 is that the maximum per-
formance enhancement of a MOE approach occurs when
(empty set). An example is to
designate each classifier to be responsible for classifying
a particular class. The assumption that
is
essential in this theorem. If
(interval between
zero and one), it is possible to find a counter example. Let
, , and
. Then .If
, then which yields correct
classification.
On the other hand, whether
takes binary values or
not, if both classifiers make correct classification, so will the
combined classifier.
Theorem 2: With the same definitions as in Theorem 1, and
(9)
Proof: Assume
[class , and ,
. Then
(10)
Thus, the output
is correctly classified.
From Theorem 2, it is clear that if both classifiers #1 and #2
correctly classify a pattern
, then the combined classifier will
also correctly classify the same pattern. Hence, this pattern can
be excluded from the user-adaptation training set as it will not
affect the result.
Adaptation Algorithm: Based on the result indicated in
Theorems 1 and 2, the design objective of the MOE network in
(3) is to devise a training algorithm to estimate the parameter
vectors
. Given that and are fixed
classifiers, this problem can be solved by a gradient procedure
as follows: Let us assume
be a set
of training data used for searching the optimal gating functions
and , such that the square error at the output
is minimized.
A gradient search algorithm can be devised as follows:
(11)
The initial values of
and are set to be the centroids of
the regions
and , respectively, for in the
user-specific training data set. The gradient of
with respect
to
can be calculated as
(12)
(13)
where
. In (13), we
assumed the transfer function
is a differentiable threshold
function, and is applied to the vector, element by element.
Finally, with (13), we have as shown in (14) at the bottom of
the page. Hence, for
, we have
(15)
Note that in above derivation, the error
is accumulated
over the entire epoch (
feature vectors). The summation
over
may be removed if we use on-line update of ’s
for each sample. This yields the following expression for
:
diag
(16)
Clearly, we have
. This is not surprising
with two parameter vectors arriving at a decision hyperplane
.
Until now, we have assumed that the user-specific ECG
beat classifier
is readily available. However, in reality
it needs to be trained with the user-specific training data set.
Also, the combined classifier
needs to be trained by
the same data set in order to determine the gating network
parameters. Therefore, if
is trained to 100% accuracy
(14)
HU et al.: PATIENT-ADAPTABLE ECG BEAT CLASSIFIER 895
on the user-specific data set, then the gating network of choice
may be
and . In light of the results
of Theorems 1 and 2, we devised the following strategy to
alleviate this problem: First, we construct the user-specific
training data set to contain only those feature vectors which
the original classifier misclassified. We further partition this
training data set into two subsets: one for the training of the
user-specific classifier
, and the other for estimating
and .
IV. E
XPERIMENT
The purpose of this experiment is to demonstrate the useful-
ness of the proposed user-adaptation procedure. In particular,
we will show that an ECG beat classifier trained on general
patient records does not perform well when presented with
patient records which contain rare beat types. Moreover, we
show that the performance of the MOE classifier is able to gain
significant performance enhancement with a small amount of
annotated patient specific training data.
A. Data Preparation
In this study, we concentrate on the classification of ven-
tricular ectopic beats (VEB’s). The 48 records (tapes) from
MIT/BIH ECG arrhythmia database [17], [19] are used for the
development and evaluation of the classifier. The availability
of annotated MIT/BIH database has enabled the evaluation of
performance of the proposed beat classification algorithm. The
American Association of Medical Instrumentation (AAMI)-
recommended practice [18] has provided a protocol for a
reproducible test with realistic clinical requirements, empha-
sizing tape-by-tape presentation of results that estimate an
algorithm’s ability to detect events of clinical significance.
Accompanying each tape in the MIT/BIH database is an
annotation file in which each ECG beat has been identified
by expert cardiologist annotators. These labels are referred to
as “truth” annotations and are used in training (developing)
the classifiers and also to evaluate the performance of the
classifiers (experts) in testing phase. According to the AAMI-
recommended practice, records containing the paced beats
(four records) can be excluded from the reporting require-
ments. Since this study is to evaluate the performance of a
classifier that can identify a premature ventricular contraction
(PVC), certain records in the database with no PVC’s (11
records) were excluded from the study, leaving 33 records of
interest. These excluded records are listed in Table I. Data
from channel 1, down-sampled to 180 samples/s were used in
this study. The selected files consist of 13 records (numbered
from 100–124, inclusive, with some numbers missing) and
20 records (numbered from 200–234, inclusive, with some
numbers missing). The first group is intended to serve as a
representative sample of a variety of waveforms and artifacts
which an arrhythmia detector might encounter in routine
clinical use. Records in the second group include complex
ventricular, junctional, and supraventricular arrhythmias and
conduction abnormalities. Several of these records are ex-
pected to present significant difficulty to arrhythmia detectors
because of the features of the rhythm, QRS morphology
TABLE I
R
ECORDS OF MIT/BIH DATABASE THAT WERE EXCLUDED FROM THE STUDY
TABLE II
F
OUR CATEGORIES OF INTEREST INTO WHICH THE
ECG BEATS OF THIS STUDY ARE CLASSIFIED
variation, and signal quality. These records were reported to
have gained considerable notoriety among database users [18].
In this experiment, we use the first group of files as the
training data to develop a GE classifier which is able to
classify typical ECG beats. The second group of 20 records
is used to simulate the ECG records of 20 patients, which
are to be classified by the GE classifier. Since these records
consist of less-frequently seen beats, it is expected that the
GE classifier will not perform well. If this GE classifier were
a commercial device, it will be deemed not-applicable (due to
low performance) to many of these 20 test records. However,
with the MOE approach, we will adapt this GE classifier with
a LE classifier to gain significant performance enhancement
at low cost.
The beats in the MIT/BIH database are of several different
types. In this study, we are interested in identifying four
different categories, as indicated in Table II. Each of the
four categories included beats of several types as shown in
Table III. The AAMI convention was used to combine the
beats into four classes of interest.
B. Training and Testing Procedure
In this study, a GE classifier was developed with SOM and
LVQ algorithms using the data from the records of the first
group (100–124). Before testing the records, a LE classifier
was developed for each of the records in the second group
using the first 2.5 min of data. The rest of the record is
then tested using the mixture of global and LE’s as explained
before. Since each record in the MIT/BIH database is of
length 30 min, the 2.5 min segment account for 1/12th of total
available patient specific data and contains approximately 150
ECG beats. In practice, the attending cardiologist or any expert
in ECG beat annotation will have to annotate a brief segment
of patient-specific ECG in order to take advantage of the
MOE approach. We believe that this is a reasonably small cost
compared to the potential gain in performance enhancement.
In future, we will explore a more effective method to further
reduce the amount of required annotated patient-specific data.