You are on page 1of 9


5, MAY 1998


Artificial Neural Networks

for Automatic ECG Analysis
Rosaria Silipo and Carlo Marchesi, Member, IEEE

Abstract The analysis of the ECG can benefit from the

wide availability of computing technology as far as features and
performances as well. This paper presents some results achieved
by carrying out the classification tasks of a possible equipment
integrating the most common features of the ECG analysis:
arrhythmia, myocardial ischemia, chronic alterations. Several
ANN architectures are implemented, tested, and compared with
competing alternatives. Approach, structure, and learning algorithm of ANN were designed according to the features of each
particular classification task. The trade-off between the time
consuming training of ANNs and their performances is also
explored. Data pre- and post-processing efforts on the system
performance were critically tested. These efforts crucial role on
the reduction of the input space dimensions, on a more significant
description of the input features, and on improving new or
ambiguous event processing has been also documented. Finally,
the algorithm assessment was done on data coming from all the
currently available ECG databases.
Index TermsArrhythmia classification, ECG, ischemia detection, post-processing, pre-processing, recurrent neural networks,
static neural networks.


LECTROCARDIOGRAPHY has a basic role in cardiology since it consists of effective, simple, noninvasive,
low-cost procedures for the diagnosis of cardiovascular disorders that have a high epidemiologic incidence and are
very relevant for their impact on patient life and on social costs. Even when some imaging techniques, such as
echocardiography or tallium scintigraphy offer more specific
diagnostic evidence in some instances, particularly related to
the reduction of myocardial perfusion, the above-mentioned
unique attributes reserve a major role to electrocardiography
in the diagnosis of cardiovascular diseases.
Pathological alterations observable by Electrocardiography
can be divided into three main areas:
1) cardiac rhythm disturbances (or arrhythmia);
2) dysfunction of myocardial blood perfusion (or cardiac
3) chronic alteration of the mechanical structure of the heart
(for instance left ventricular hypertrophy).
Manuscript received July 23, 1997; revised October 13, 1997. The associate
editor coordinating the review of this paper and approving it for publication
was Prof. Yu-Hen Hu.
R. Silipo is with the Department Sistemi e Informatica, Universit`a di
Firenze, Firenze, Italy (e-mail:
C. Marchesi is with the Department Sistemi e Informatica, Universit`a di
Firenze, Firenze, Italy. He is also with the Istituto di Fisiologia Clinica (CNR),
Pisa, Italy (e-mail:
Publisher Item Identifier S 1053-587X(98)03179-1.

An indirect indication of the importance of these disturbances

is reflected by the availability of only three types of independent databases for performance evaluation of the algorithms
for ECG analysis. They consist of archives of annotated signals
representative of arrhythmia, ischemia, and chronic diseases.
Moreover, the available equipment implement separately only
one specific procedure at a time out of these areas of cardiac
Presently, many factors seem to be combined and give new
momentum to research and development in electrocardiography:
Ambulatory monitoring of the electrocardiogram (ECG)
has established its role in many circumstances, particularly as a detector of rare episodes and for therapeutic
control. Such circumstances, occurring in the various
classes of myocardial ischemia and in arrhythmia, have
shown the improvement derived from such a dynamical
The increasing availability of low-cost high-performance
computing technology encourages improvement in electrocardiography by offering a reliable and comprehensive
solution to the automatic diagnosis of the ECG.
The rising cost of health care asks basically for a reduction in the circumstances when the admission of patients
to hospitals is recommended. Here, a new approach to the
design of portable devices for ambulant subjects can help
in making the patient less hospital dependent.
Thus, designing low-cost, high-performance, simple to use,
and portable equipment for electrocardiography, offering a
combination of diagnostic features, seems to be a goal that
is highly worthwhile. Such equipment should embed and
integrate several techniques of data analysis, such as signal
processing, pattern detection and recognition, decision support,
and human computer interaction.
The literature in this topic reports several approaches to
classification, including Bayesian [1] and heuristic approaches
[2], expert systems [3], and Markov models [4]. In general,
past approaches, according to published results, seem to suffer
from common drawbacks that depend on high sensitivity to
noise and unreliability in dealing with new or ambiguous
Artificial neural networks (ANNs) have often been proposed as tools for realizing classifiers that are able to deal
even with nonlinear discrimination between classes and to
accept incomplete or ambiguous input patterns. Recently, the
connectionist approach has also been applied to ECG analysis
with promising results [5][9].

1053587X/98$10.00 1998 IEEE


Despite so many applications of ANN to ECG analysis,

the lack of a complete, systematic, and exhaustive comparison
with traditional methods currently implemented in ECG analyzers does not allow a more organized industrial planning of
ANN applications as ECG analyzers. Based on the current
state of the art, it is not clear to what extent ANNs can
answer the two goals listed above. Do they offer a more
reliable automatic analysis of the ECG signal? If yes, in
what fields? The development of a reliable automatic ECG
analyzer covering the most important areas of ECG analysis
would produce a tool able to reduce medical dependence in
low-risk conditions, like in patients with moderate ischemia
or arrhythmia or for early diagnosis through such a selfmonitoring technique. In this work, the role of ANN has
been widely explored in order to define the procedures for
ECG pattern classification that are well suited to carry out
a reliable engine for equipment integrating such features like
arrhythmia classification, ischemia detection, and recognition
of chronic myocardial diseases. Particular care has been given
to preprocessing techniques because the quality of the signal
entering the classifier is an essential determinant of the final
quality of the classification.
As far as the quality of the performances, we are faced
with a basic problem; a theory of electrocardiography is
lacking, that is, no definite evidence may be claimed about the
correspondence between an ECG sign in a particular subject
and its pathophysiological meaning. The only reference we
have to assess the quality of equipment is the recommendations
of specific authorities, generally inviting them to evaluate the
systems on widely available databases. Thus, we have validated our algorithms on the above-mentioned ECG databases
[1], [10], [11].
The paper presents the three modules we have developed
as well as their validations. The differences of the methods
to be used for carrying out the classifiers suggest that we
deal separately with the three topics, which are summarized
as follows.
An ANN structured as an autoassociator is implemented
to perform arrhythmia detection because of its capability to reject unknown or ambiguous patterns. For
this purpose, two uncertainty criteria are introduced and
Both a static and a recurrent ANN approach are implemented in several architectures to detect ischemic
episodes. While the first approach features an easier
learning process, the second one is able to learn the
input signal evolution even on a reduced training set. The
trade off between performance and computational cost is
investigated for both approaches.
Recognition of chronic myocardial diseases requires a
three step procedure. The parameters chosen for analyzing
the ECG have been fuzzy processed by a layer of normalized radial basis functions (NRBFs) and then have been
analyzed by a neural network; finally, a pruning technique
is applied to reduce the network size and to improve its
generalization capability.


Fig. 1.

ECG waveform.

Since the evaluation methods have the same background in

all the modules, a paragraph on some aspect of ECG signal
analysis and testing introduces general definitions and criteria.
An ECG signal is the manifestation on the body surface of
the myocardium electrical activity, which appears as an almost
periodic signal. Each cardiac pathology affects permanently or
temporarily, completely or partially, the ECG basic waveform
(Fig. 1).
A set of algorithms for signal conditioning, QRS complex
detection (Fig. 1), delineation and measurement of wave
amplitude, duration, and area are usually adopted to perform
the appropriate parameter extraction for the proposed ECG
interpretation task. The cardiac pathology whose presence can
be assessed by means of ECG analysis consists mainly of
arrhythmia, ischemia, and some chronic cardiac diseases. To
these groups of diseases correspond three types of independent
ECG databases, which are developed to allow a standard
approach of performance evaluation of the algorithms for
analysis of
1) arrhythmias (developed at MIT-Beth Israel Hospital
[MIT-BIH], Boston, MA);
2) ischemia-induced changes of ECG waveforms (developed through a project supported by the European Union
and by the European Society of Cardiology, lead by the
CNR Institute of Clinical Physiology, Pisa, Italy);
3) chronic alterations (developed through the CSE project
and supported by the European Union and lead by the
University of Leuven, Leuven, Belgium) [1], [10], [11].
For the comparison of traditional with ANN-based classifiers in arrhythmia classification, 30 out of the 35 records
of a part of the European database (called VALE) are used.
The VALE database consists of single-channel ECG records
lasting 3 h and sampled at 200 samples/s [10]. In addition, 16
two-channel 30-min ECG records sampled at 360 samples/s
are extracted from the MIT-BIH database [11] in order to
investigate ANNs uncertainty management in case of new or
ambiguous events.
The systematic comparison among the static and the recurrent approaches in detecting ischemic episodes has been


performed on the European ST-T database [10]. It contains

100 two-channel ECG records lasting 2 h and sampled at 250
To evaluate the ANNs performance in recognition of
chronic cardiac diseases, the CSE database has been used
[1]. Here, the ECG signal is characterized by 37 average
primary measurements plus the age and gender of the patient.
The system performance is usually quantified by means of
two parameters: sensitivity and specificity or sensitivity and
positive predictive accuracy (PPA), when the definition of false
negatives may be questionable.
indicates the rate of true positive events
measures the rate of true negative
for class ,
is the rate of the true positive
events for class , and
events among all the classified events in class .
number of diagnostic classes;
number of events
referred to the class by the
process , with being.

is the event
assigned to class ;

is the event
belongs to class ;

is the event
is assigned to the uncertain class.


QRS complex. In fact, correct detection of the QRS is the

key to reliable measurement of the parameters to be used for
further analysis [7]. Our QRS detector is implemented by using
the spatial velocity as the evidence variable, according to the
AHA recommendations [12]. It also includes noise estimation
criteria. The detector has been validated on the MIT-BIH and
ST-T databases; its sensitivity and specificity was very close
to 100% [13].
The most frequent forms of arrhythmia are ectopic beats of
a ventricular or atrial (supraventricular) origin. A premature
occurrence is a common attribute of them. Ventricular ectopic
beats exhibit a different shape from sinus rhythm, whereas the
shape of supraventricular ectopic beats resembles the normal
one. Thus, the samples of each beat and a measure of its
are taken into account as input
prematurity degree
The sample input sequence starts at the beginning of the
QRS complex and includes a 150-ms time window, where
the beat morphology is supposed to be adequately described.
for the beat is defined as

The percentage of uncertain beats, when uncertain beats can
be defined, is calculated as


Sometimes, a global error index is useful to resume

. Since this statistical index requires
equally distributed classes, some weights
are recursively
defined as

and the error index

is defined as


A feature that is common to the modules and vital to the

quality of the global system is the synchronization with the
cardiac beat, which is obtained through the detection of the

represents the
interval before the current
beat (Fig. 1), and is set to 20; the
of the arrhythmic
beats are excluded from (7).
The inclusion of arrhythmic beats in (7) does not dramatically change the performance of either the ANN system or of
the traditional classifier but affects it in the presence of runs
of such arrhythmic beats. In this case, the reference condition
would be completely changed, and the system would be fed
with incorrect information.
The randomness of the occurrence of arrhythmic beats
suggests that a static analysis, based only on the features of
the current beat, might be appropriate. The high intra and
interpatient variability of the beat shape suggests an approach
that takes the patient as a reference of himself or herself.
Therefore, the first 15% of each ECG record is used to build
the training set, and the remaining part is used for testing.
A common drawback of arrhythmia classifiers is their trend
to misclassify new or ambiguous beats. Among the possible
alternatives [7], the structure of the arrhythmia detector that
resulted best suited to solve the problem was an autoassociator
because of its intrinsic capability to reject unknown patterns
[14]. It is trained with the back propagation (BP) algorithm
[15] to reproduce the beat shape as the output pattern. A class
code is assigned to each beat and is added as a supplementary
part to the output layer (Fig. 2).
The very desirable property of rejecting ambiguous ECG
beats when the output pattern is not recognized as a proper
beat is then achieved according to two uncertainty criteria.
The first one rejects all the classes that are not represented
within the training set and have a consistent morphology,




Fig. 2. Autoassociator for arrhythmia classification.

according to the condition


uncertain class

number of the signal samples;
value of the input unit ;
value of the corresponding output unit
for the
The second uncertainty criterion deals with ectopic beats
with pseudo-normal shape. The class showing a maximum
score is chosen only if it is not uncertain, that is, if the
difference between the two output values with the highest
score does not satisfy


uncertain class

denotes the output class for pattern .

After several experiments,
has been set to 0.1,
has been set to 0.6.
The first uncertainty criterion has been shown to be successful in rejecting beats not included in the training set and
having a characteristic morphology, such as the ventricular
ectopic beats (V), junctional beats (J), ventricular escape beats
(p), some cases of aberrated beats (a), and others. For example,
when a training set with only normal (N) and supraventricular
ectopic beats (S) is used, the V beats will be rejected (Table I).
The second criterion is less effective so that only very
premature S beats are rejected. For example, only 7% of S
beats are refused when the training set is composed of V and
N beats (Table I). The effect of each one of the two uncertainty
criteria used separately is shown in Table I for different
compositions of the training set: N, S , N, V , and N, S,
V . An example with a training set that is comprehensive of
all the most important output classes is also reported (Table I).
A comparison with a more traditional arrhythmia classifier
has been performed without using uncertainty management
(Fig. 3). A clustering method based on the median algorithm
is adopted because it produced the best classification rates

Fig. 3.

ANN versus clustering method.

out of a group of the most used arrhythmia classifiers [7].

Perhaps the reason for the better performance of the ANNs
classifier versus the clustering method consists of its capability
to learn complex separation surfaces, which are required in
supraventricular arrhythmic beat discrimination. Moreover, the
clustering method needs additional work to define logical
rules to analyze beat prematurity [7], whereas ANNs self
learning allows an implicit and more accurate definition of
such decision rules that are implemented by appropriately
Ischemic episodes appear as slow changes in the ST segment
(Fig. 1). Artifacts can also be encountered due to movements
of the body position related to electrode recording position.
They are characterized by ST sudden changes and limit the
reliability of every automatic ST change detector. Further
limiting factors, like the slow trend of the ST changes and
erratic specificity of the usual ST parameters, are still waiting
for a reliable solution.
The nonlinear ANNs processing might be able to discriminate the ST changes related with myocardial ischemia,
but the large variety of topological architectures and learning
paradigms makes the choice hard. On one side, the static
approach leads to systems with reduced dimensions and with
a fast learning process. On the other side, recurrent neural
networks (RNN) are supposed to be able to capture data
dynamics. In this paper, the static and the dynamic ANNs


approaches are compared against each other and against conventional methods.
The traditional ST analysis applies empirical thresholds in
amplitude and in time to the ST parameter time series [13].
To improve such an analyzer, ST amplitude and slope are first
sent to a preprocessing unit and carried out by a static ANN
(sANN) trained with BP using ST elevations, ST depressions,
normal patterns, and artifacts. However, the introduction of
the ANN unit does not show significant improvement (Fig.
5) since the drawback of disregarding time course dominates
over the benefits of the nonlinear feature emphasis.
To associate time information into the input data, sequences
of frames
are defined on the data. The frame
consists of the samples in a 400-ms window on the ST
segment of the beat .
A dimension-reduction technique based on the principal
component analysis (PCA) transform is applied to the ST
segment. A suboptimal approach is adopted because it allows the time-consuming eigenanalysis required by PCA. A
representative training set of 97 663 waveforms, extracted
from 105 15-min ECG records, is used to build the ECG
autocorrelation matrix, and the eigenvectors required by the
PCA are calculated just once instead of cycle by cycle. It was
found that the first four PCs represent about 90% of the ST
segment energy [16].
A median filter is subsequently applied to the PC time series
samples, the
to reject outliers. Given a time window of
median filter substitutes its central sample with the median
value among the average of the first samples, the central
sample, and the average of the last samples of the window.
A two-layer feedforward static ANN is trained using the BP
algorithm with sequences of seven frames of four PCs of each
ST segment (PCA sANN). Such dimension reduction of the
input space allows training of the designed sANN with a more
comprehensive data set, including several artifact examples.
RNNs represent the easiest extension of ANNs to the analysis of sequences because of their capability to capture data
dynamics. Sequences of four ST segments are employed as an
input vector. The forwardbackward training algorithm [17]
is adopted. A recurrent layer with auto and left crossconnections is introduced in the previous sANN architecture (RNN).
Autoconnections are assumed to be necessary for information
latching, whereas crossconnections, which are progressively
activating hidden neurons, are supposed to follow the ST
segment pathological evolution. After several experiments
[20], the number of hidden units is chosen to be 40 and that
of the crossconnections toward left units to be 3. Because of
the high number of free parameters, the training can only be
performed after excluding artifacts from the training set.
Embedding some a priori knowledge into the structure,
that is, implementing a so called knowledge-learning network
(KLN) [18], can help in reducing the number of free parameters. The goal consists of modeling the RNN structure in such
a way that only a progressive change in ST segment excites
the output neurons. That might characterize the dynamic of
pathological ST changes versus artifacts.
In order to simplify the process of a priori knowledge
translation, the average value of the ST segment is calculated


Fig. 4. KLN structure.

by an input neuron with fixed weights. A layer of NRBF

units added to the previous RNN structure codes the state of
this average value (Fig. 4). Each NRBF covers a different
interval of the ST amplitude range. The NRBF units with
similar central values are placed topologically close to each
other so that for smooth ST level changes, the input sequence
progressively excites adjacent NRBF units. In the case of
artifacts, the excited NRBFs are topologically far from each
in the
The following rules are defined for each unit
upcoming recurrent layer :

where the input and the output of every neuron are supposed
to be in the range of [0, 1],
denotes the output of the
at the time , and
denotes the corresponding
input at the time . Such rules lead to the constraint equation
represents the threshold value of the activation
functions of the neuron
, and
represents the
connection weight between the neurons
in the layer and
in the layer .
The number of crossconnections in the recurrent layer is a
critical parameter to be carefully chosen, especially when the
number of NRBF units is high. Here, the number of NRBF
units is fixed to be 21. Due to the large number of constraints,
the training process cannot be extensive, and the KLNs results
are very sensitive to noise.
A criterion about the minimum duration of an ischemic
episode is always applied to the system output, as defined
in [10] and [19]. The recurrent approach, while effective in
capturing dynamics out of the input data, is globally difficult
to be interpreted and time consuming. The results of static
versus recurrent approach comparison are shown in Fig. 5. The
combination of PCA and sANN produces sensitivity and PPA
values comparable with those of the recurrent approach, but




Fig. 5.

Static versus recurrent approach.

the PCA sANN approach requires a faster and easier training

process (Fig. 5).
Such an algorithm shows a higher PPA value. In the
literature, the usual PPA is found to be around 8082% [10],
[19], while the sensitivity can vary from 70 to 80% after an
accurate setting of the thresholds.
Since not all the examples of artifacts can be included into
the training set, the static approach fails to recognize episodes
with an atypical time evolution pattern. The recurrent approach, on the contrary, is trained to classify only slow or fast
time evolution, and its misrecognition of ischemic episodes
is mainly relative to noisy or weak episodes. The second
episode in record e0106 in the ST-T European Database can
be reported as an example. This episode is particularly hard
to recognize since it is placed immediately after an artifact. In
this case, the RNN structure recognizes the ischemic episode,
whereas PCA sANN fails because of the unusual pattern
configuration. Threshold-based algorithms fail too because the
reference baseline changes too fast to allow the algorithm to
adapt its parameters to the new setting.
Myocardial diseases, due to chronic electrical or anatomical
problems, produce permanent changes in the ECG waveform
(Fig. 1). The most common diagnostic classes investigated
with the ECG are referred to as
left ventricular hypertrophy (LVH);
right ventricular hypertrophy (RVH);
bilateral ventricular hypertrophy (BVH);
inferior myocardial infarction (IMI);
anterior myocardial infarction (AMI);
mixed myocardial infarction (MIX);
besides, of course, the normal status.
The ECG analysis is performed on 37 average simple and
composite ECG measurements plus some data from patient
history, such as age and gender. A random set of 2446 patients
has been selected from the CSE database [1] for the learning
phase of the ANNs, and the remaining 820 cases have been
used for testing, i.e., evaluation of the system performance.

Since the input parameters result from an average process,

they do not carry defined information but can point to overlapped diagnostic classes. Thus, a level of linguistic description that is able to describe these imprecise medical concepts
is adopted to preprocess the data and is performed by applying
a layer of normalized radial basis function (NRBF) units [21].
For each input parameter , seven NRBFs are defined (one
for each output class), with


is the number of diagnostic classes, and

are the central value and the dispersion factor of the bell. The default values of
shaped function NRBF
of the unit NRBF
are derived from the statistics of the
training set, i.e., the estimated mean and standard deviation of
in the diagnostic class .
the input parameter
The outputs of the NRBF nodes represent the membership
degrees to some linguistic terms and are fed into a classical
feed-forward two-layer sigmoidal neural network (Fig. 6).
The resulting hybrid system is considered to be in the same
connectionist framework, and the BP algorithm is modified to
allow the training of all the parameters. Because of the system
high dimension, a pruning procedure is applied by adding a
penalty term to the training cost function [21].
In order to analyze the possible improvements derived from
the NRBF and/or input weight training, all the combinations
for training have been evaluated (Table II). The results show
that the modified BP algorithm is able to build good NRBF
if the original ones are not accurate enough. In this way, the
critical point of the choice of the most adequate preprocessing
functions for data abstraction can be avoided. Such a detailed
linguistic description of the input space makes the task of
the upcoming neural network easier since it represents almost
all the diagnostic knowledge in a way that is close to the
physicians reasoning. Moreover, the combination of the two
learning processes leads to a more representative abstraction
of the input data.
The average sensitivity and specificity that is calculated first
for each class over all the database cases and then over all
the diagnostic classes, are used to characterize the system
The NRBF preprocessing and the modified BP algorithm
effectively improve the performance of the final system. In
Table II, the classification task is shown to be more reliably
performed when the NRBF preprocessing is applied and
the NRBF parameters are trained. The pruning procedure



Fig. 6. NRBF architecture.

improved the performance of 2% in sensitivity and allowed

a reduction of 37% of the network weights and of 24% of the
network nodes when both NRBF parameters and weights are
trained. The obtained results are comparable with those from
the best-known ECG classification algorithms [1].
A. ANN vs. Traditional Classifiers
Looking globally at the diagnostic performance issue, it may
be stated that ANNs offer a promising alternative to current
The ANN structured as an autoassociator exhibits better performance than traditional clustering algorithms in arrhythmia
classification. Its capability of building nonlinear separation
surfaces in the input space results in a more reliable classification of supraventricular arrhythmia (Fig. 3). The autoassociator
learns autonomously how to weight the beat prematurity
measure. This is in contrast with the clustering method, where
the definition of the corresponding logic rules is more time
The comparison between the recurrent and static ANNs
approach in ischemia detection does not show appreciable
difference in terms of sensitivity and PPA, even though the
static approach should be preferred because of its faster
training process. The missing property of learning signal
dynamics is balanced by the chance of a more representative
training set.
Comparing the neural approach with the traditional
threshold-based algorithm, we can notice an improvement
in the number of rejected artifacts as the PPA percentage
indicates (Fig. 5). In addition, it has to be remarked that the

performance of the threshold criterion is strongly dependent on

the quality and the reliability of the previous QRS detection.
The ANNs systems, on the contrary, analyze the whole
ST segment, reducing in that way such dependence on the
reliability of the QRS detector.
All the ANN-based ischemia detectors reach the performance of an expert [10] and are able to recognize ST episodes
with such a low amplitude that they are not even annotated in
the database. It is interesting to note that the medical expertise
translated into the neural structure produces a bad copy of the
traditional rule-based ischemia detectors. Therefore, a set of
a few free parameters left for training leads to a final KLN
that is too sensitive to noise and unable to generalize to new
examples even if they are slightly different from the theoretical
ST changes.
The neuro-linguistic system adopted for recognition of
chronic myocardial diseases reaches performances that are
comparable with those of traditional ECG classifiers. This is
basically due to the low diagnostic information content of the
ECG for detecting chronic pathologies since infarctions and
hypertrophies, especially the ones with old origin, can show
up with a nonstandard ECG wave-shape depending on the
patients clinical history.

B. Post- and Pre-processing Techniques

Due to the biological nature of the ECG signal, the case
of new or unknown events is quite frequent and has to be
taken into account even more when the analysis is patientdependent, as in the case of arrhythmia analysis. In this case,
it is impossible to find examples of all the possible diagnostic
classes in the first minutes of the patient record. Then, the



probability of finding new or ambiguous events during the

analysis increases dramatically.
The proposed combination of the ANN approach with
uncertainty criteria enhances the capability of uncertainty
management of the neural approach. These uncertainty criteria
lead to a very high rejection rate of unknown or ambiguous
events and very good performances with a global error close to
0 (Table I). Even though this is possible, it would be useless to
include all the ECG arrhythmic classes in the training set since
the consequent high dimension would not allow the training
process to converge to an adequate error value. According to
the high reliability of the uncertainty management proposed
for arrhythmia analysis, it would be more convenient to define
some main diagnostic classes and to analyze the uncertain
beats by means of differently trained subsystems.
Not all the reduction techniques are useful to reduce the
redundancy of the input data because of possible cancelling
of diagnostic information. For example, the minimum mean
squared error criterion adopted in the signal reconstruction
does not guarantee savings of low-energy ECG waves potentially with high discriminating power. For this reason the PCA
adopted in the ischemia detector with satisfactory results is not
suitable to other classification tasks since it optimizes mean
squared error while disregarding potentially useful diagnostic
The choice of an adequate preprocessing technique represents a crucial point in every ECG analyzer design but
even more in ANNs systems. A very accurate description of
the input space makes all classification much more accurate.
This makes the training process easier and improves system
The relevance of an accurate preprocessing is obvious
when designing the ischemia detector. The high dimension
of the input space and the large variety of ischemic episodes
and artifacts required a dramatic reduction of the problem
dimension. The PCA allowed reduction of each input vector
from 40 components to the first four PCs and then to train
even a static ANNs structure with sequences of data from a
very extensive training set.
In the recognition of chronic myocardial diseases, the NRBF
linguistic preprocessing allows a more detailed description of
the input features, an easier and faster learning process, and,
finally, a possible interpretation of the network decisions.
A final comment should be made concerning real-time
use of portable equipment. In this case, the training may be
inaccurate or very complex because of the unpredictability
of the real-time events. Solutions allowing adaptation of the
system parameters should be devised.
Several ANNs structures combined with different preprocessing and postprocessing techniques are designed and
evaluated for arrhythmia classification, ischemia detection, and
recognition of chronic myocardial diseases. The capability of
uncertainty management of an ANN structured as an autoassociator is deeply investigated in arrhythmia classification. Static
and recurrent ANN approaches are implemented and compared

in several architectures for detecting ST changes related to

myocardial ischemia. The role of different techniques to reduce
the input dimension is discussed. Linguistic data preprocessing
is introduced for recognizing chronic myocardial diseases. A
pruning technique is applied to reduce the system dimension
and to improve its generalization capability as well.
Every ANN has been tested and compared with the most
common traditional ECG analyzers on appropriate databases.
Thus, based on the results, the ANNs approach is shown to
be capable of dealing with the ambiguous nature of the ECG
The crucial role of data preprocessing and postprocessing
comes out, either for reducing the input space dimension or
for more appropriately describing the input features.

[1] J. L. Willems and E. Lesaffre, Comparison of multigroup logisitic and
linear discriminant ECG and VCG classification, J. Electrocardiol.,
vol. 20, pp. 8392, 1987.
[2] J. L. Talmon, Pattern Recognition of the ECG. Berlin, Germany:
Akademisch Proefscrift, 1983.
[3] A. Gallin et al., A computer system for analysis of ST segment changes
on 24 hours Holter monitortapes: Comparison with other available
systems, J. Amer. Coll. Cardiol., vol. 4, pp. 245252, 1984.
[4] A. D. Coast, R. M. Stern, G. G. Cano, and S. A. Briller, An approach to
cardiac arrhythmia analysis using hidden Markov models, IEEE Trans.
Biomed. Eng., vol. 37, pp. 826835, 1990.
[5] L. Edenbrandt, B. Devine, and P. W. Macfarlane, Neural networks for
classification of ECG ST-T segments, J. Electrocardiol., vol. 25, no.
3, pp. 167173, 1992.
[6] T. Stamkopulos, M. Strintzis, C. Pappas, and N. Maglaveras, One lead
ischemia detection using a new back-propagation algorithm and the
European ST-T database, IEEE Comput. Cardiol., pp. 663666, 1992.
[7] R. Silipo, M. Gori, A. Taddei, M. Varanini, and C. Marchesi, Classification of arrhythmic events in ambulatory ECG, using artificial neural
networks, Comput. Biomed. Res., vol. 28, pp. 305318, 1995.
[8] J. H. Frenster, Neural networks for pattern recognition in medical
diagnosis, in Proc. 12th IEEE EMBS Biomed. Eng. Perspectives: Health
Care Technol. 1990s Beyond, P. C. Pedersen and B. Onaral, Eds., 1990,
pp. 14231424.
[9] G. Bortolan and J. L. Willems, Diagnostic ECG classification based on
neural networks, J. Electrocardiol., vol. 26, pp. 7579, 1994.
[10] A. Taddei et al., The European ST-T database: Standard for evaluating
systems for the analysis of ST-T changes in ambulatory electrocardiography, Euro. Heart J., vol. 13, pp. 11641172, 1992.
[11] MIT-BIH database distributor, Tech. Rep., Beth Israel Hospital,
Biomed. Eng., Division KB-26, Boston, MA.
[12] J. J. Bailey et al., Recommendations for standardization and specification in automated electrocardiography: Bandwidth and digital signal
processing, Circulation, vol. 81, 1990.
[13] R. Silipo, A. Taddei, and C. Marchesi, Continuous monitoring and
detection of ST-T changes in ischemic patients, IEEE Comput. Cardiol.,
pp. 225228, 1994.
[14] M. Bianchini, P. Frasconi, and M. Gori, Learning in multilayered
networks used as autoassociators, IEEE Trans. Neural Networks, vol.
6, pp. 512515, 1995.
[15] D. E. Rumelhart, G. E. Hinton, and R. J. Williams, Learning internal
representations by error propagation, in Parallel Distributed Processing. Exploration in Microstructure of Cognition, J. A. Feldman et al.,
Eds. Cambridge, MA: MIT Press, 1986.
[16] R. Silipo, P. Laguna, C. Marchesi, and R. G. Mark, KarhunenLo`eve
transform and artificial neural networks for ST-T analysis, IEEE
Comput. Cardiol., pp. 213216, 1995.
[17] B. A. Pearlmutter, Learning space trajectories in recurrent neural
networks, Neural Comput., vol. 1, no. 2, pp. 263269, 1989.
[18] P. Frasconi, M. Gori, M. Maggini, and G. Soda, Unified integration
of explicit knowledge and learning by example in recurrent networks,
IEEE Trans. Knowl. Data Eng., vol. 7, pp. 340346, July 1995.
[19] F. Jager, G. B. Moody, S. Divjak, and R. G. Mark, Assessing the
robustness of algorithms for detecting transient ischemic ST segment
changes, in IEEE Comput. Cardiol., pp. 229232, 1994.


[20] G. Bortolan, R. Degani, and J. Willems, ECG classification with neural

networks and cluster analysis, IEEE Comput. Cardiol., pp. 177180,
[21] F. Piantini, R. Silipo, C. Marchesi, and M. Gori, Recurrent neural
network for electrocardiogram analysis, in Proc. ICANN, 1994, pp.
[22] G. Bortolan, R. Silipo, and C. Marchesi, Fuzzy pattern classification
and the connectionist approach, Pattern Recognit. Lett., vol. 17, pp.
661670, 1996.

Rosaria Silipo received the electrical engineering

degree from the University of Florence, Florence,
Italy, in 1992 and the Doctorate degree in bioengineering from the Politecnico di Milano, Milan, Italy,
in 1996.
She was a Visiting Researcher at the Massachusetts Institute of Technology, Cambridge, in 1994
and with Siemens AG, Munich, Germany, from
1996 to 1997. She is currently an External Fellow at
the International Computer Science Institute (ICSI),
associated with the University of California, Berkeley. During her Ph.D. studies, she investigated the different fields of automatic
ECG analysis. Her current research and interests include brain modeling based
on EEG and MEG.


Carlo Marchesi (M73) was born in Este, Padova,

Italy, in 1939. He received the degree in electrical
engineering from the University of Padova in 1967.
After a postgraduate research training at University of Pisa, Pisa, Italy, and at Duke University,
Durham, NC, he joined the Istituto di Fisiologia
Clinica (IFC), Pisa. He served as Adjunct Professor
at the University of Pavia and Padova. Since 1987,
he has been the Associate Professor of Bioengineering at the University of Florence, Florence, Italy.
He collaborates with IFC as Head of the Medical
Informatics Department and with the University of Bologna, Bologna, Italy,
in the Ph.D. program in bioengineering. His present research interests include
medical signal processing and medical decision making. He represented Italy
in several European Union meetings to plan biomedical equipment research.
Dr. Marchesi was appointed as a Member (19841993) of the Board of
Directors of Computers in Cardiology, Inc., and in 1988, he was nominated
Fellow of the European Society of Cardiology.