Author's personal copy

Medical Engineering & Physics 32 (2010) 1131–1136

Contents lists available at ScienceDirect

Medical Engineering & Physics

journal homepage: www�elsevier�com/locate/medengphy

Characterizing atypical patterns of Heart Rate before

Paroxysmal Ventricular Tachycardia
E. Álvarez a�∗ , J. Jiménez a , F. Moleiro b , A. Rodríguez b
a
Laboratorio de Fenómenos No Lineales, Escuela de Física, Facultad de Ciencias, Universidad Central de Venezuela, Caracas, Venezuela
b
Sección de Cardiología Experimental, Instituto de Medicina Tropical, Facultad de Medicina, Universidad Central de Venezuela, Caracas, Venezuela

� r t i c l e i n f o � b s t r � c t

Article history: Results of a comparative analysis between Heart Rate (HR) patterns occurring before the onset of Parox-
Received 25 January 2010 ysmal Ventricular Tachycardia (PVT) and from healthy subjects are shown. Two study groups were made
Received in revised form 6 June 2010 after electrocardiographic dynamical monitoring (Holter) of volunteers. The first group includes 100 h
Accepted 11 August 2010
from 27 healthy control subjects, and the second group consists of 88 h, ending with self-terminating
episodes of PVT from 55 patients. Patterns are defined as sequences of consecutive RR intervals, while
Keywords:
atypical patterns are defined as corresponding to unlikely behavior of Heart Rate in healthy subjects. We
Cardiac arrhythmias
investigated spatial and temporal distributions of these patterns in order to find early signs of PVT. We
Heart Rate Variability (HRV)
Holter monitoring
found that they can be grouped on a reduced number of clusters, and the number of atypical patterns
Patterns increases as we approach the onset of the episode.
Predictability © 2010 IPEM. Published by Elsevier Ltd. All rights reserved.

1� Introduction patterns, which are d-dimensional vectors whose components are

Cardiac arrhythmias are disorders of the regular rhythmic beat-
ing of the heart in which there is abnormal electrical activity.
The heart beat may be too fast or too slow, and may be reg-
ular or irregular. They can occur in a healthy heart and be of
minimal consequence. They also may indicate a serious prob-
lem and lead to heart disease, stroke, embolus or sudden cardiac
death.
Different approaches have been developed for risk stratification
of fatal cardiac arrhythmias in patients who are at risk and whose
lives might be benefited from prophylactic therapy [1–4].
Despite these achievements, transitions between rhythms of
the heart occur by different scenarios and prediction of the time
of onset of arrhythmias remains a distant goal. This is an issue
of major practical interest to the extent that predictions are more
accurate and fast, and that methods used are noninvasive and inex-
pensive. In this sense, several measures of Heart Rate Variability
(HRV) are powerful tools for the analysis of certain types of arrhyth-
mias [5–14]. However, characteristics of changes in HRV preceding
the onset of ventricular arrhythmias are unclear when using con-
ventional techniques [15–18].
As a complementary contribution to the problem of character-
ization, we focus on the analysis of the Heart Rate by studying
∗ Corresponding author. Tel.: +58 212 6051118; fax: +58 212 6051675.
E-mail addresses: ealvarez@fisica.ciens.ucv.ve, ealvarezve@gmail.com
(E. Álvarez).
defined as sequences of d consecutive RR intervals. In this way, we
are not only paying attention to the values of RR intervals, but also
the chronological order in which they occur, with the aim of inves-
tigating whether these patterns contain useful information to find
early signs of PVT.
In the next section, we introduce a simple and fast compu-
tational method to compare patterns from the HR series of two
databases: one consisting of patterns from healthy volunteers that
are a representative catalogue of the behavior of HR in this group,
and the other of patterns from records previous to PVT.
We use Euclidean distance in order to establish similarities and
differences between patterns from the two study groups. In par-
ticular, we focus on those patterns before arrhythmias which are
very different to any other pattern in the catalogue (i.e. are atypi-
cal).
In order to characterize records before PVT, we computed the
number of atypical patterns in each of them, and overall statis-
tics results are presented in Section 3. In order to explore the
potential uses of this characterization such as predicting the onset
of arrhythmias, we have investigated properties of the distribu-
tion of these patterns through the hour prior to the arrhythmia
episodes. First, we examined the spatial distribution of these
patterns in the region were they lie, and evaluated the extent
in which atypical patterns have a particular profile that occurs
repeatedly within records. Second, we analyzed their time dis-
tribution comparing the number of atypical patterns that arise
near the onset of episodes of PVT with the number that occur dis-
tantly.

1350-4533/� – see front matter © 2010 IPEM. Published by Elsevier Ltd. All rights reserved.

doi:10.1016/j.medengphy.2010.08.004
 Author's personal copy

1132 E. Álvarez et al. / Medical Engineering � Physics 32 (2010) 1131–1136

Fig� 1� Building patterns from a record of healthy subject. From the long series on the top, different 5-dimensional patterns are shown in the bottom.

In addition, we studied the differences between the evolution provide a representative catalogue of the behavior of HR in healthy
of atypical patterns in records before PVT and patterns in the cata- subjects.
logue. In order to define patterns of HR, let us first denote the records
Concluding remarks are presented in Section 4. from healthy volunteers by:

(RR)r1 � (RR)r2 � . . . � (RR)ri � (RR)rN r (1)

2� Patterns of Heart Rate
Records were obtained using a magnetic recorder (3-channels,
Rozinn Electronics, Holter Recorder Model 151) in the Section of
Experimental Cardiology of the Institute of Tropical Medicine of
the Universidad Central de Venezuela. A digitalization system was
developed and installed on a Silicon Graphics Work Station. The
digitalization was made at 500 samples per second with 8 bits of
A/D resolution and a standard algorithm for detecting the R peaks
was employed [19,20].
The generated database consists of:
A. 88 one hour records of the RR intervals ending with self-
terminating episodes of PVT from 55 patients (33 males, mean
age 58).
B. 100 one hour records of the RR intervals from 27 healthy sub-
jects. None of them had clinical symptoms nor cardiac disease
after they were evaluated according to the following protocol:
- Clinical Evaluation
- Chest X-Ray
- Echocardiogram
- Signal Averaged ECG
- Stress Test
- Standard ECG (12 derivations)
- Holter’s Dynamical Electrocardiogram (24 h)
The PVT episodes were identified by means of the analysis of
the Holter records. Besides that they finish with the onset of the
arrhythmias, all records have a low noise level and began at least
1 h after previous events (if any).
The 100 one hour records from healthy subjects were ran-
domly selected, without distinguishing day’s hour, sex or age of
the volunteers. For these reasons, we assume that these records
where the superscript r = 1, 2, . . ., 100 identifies the record and the
subscript i = 1, 2, . . ., Nr the corresponding values of the beat to beat
intervals (RR intervals).
Next, for each record we build all possible sequences of d con-
secutive RR intervals:
vr1 = (RR)r1 � (RR)r2 � (RR)r3 � . . . � (RR)rd
vr2 = (RR)r2 � (RR)r3 � (RR)r4 � . . . � (RR)rd�1
..
.
vrn = (RR)rn � (RR)rn�1 � (RR)rn�2 � . . . � (RR)rn�d�1
..
.
vrN r −d�1 = (RR)rN r −d�1 � (RR)rN r −d�2 � (RR)rN r −d�3 � . . . � (RR)rN r
thus obtaining Nr − d + 1 d-dimensional vectors that contain infor-
mation about the behavior of the HR of the rth record. These are
the vectors that we call HR patterns. This procedure is illustrated
in Fig. 1 for the case d = 5.
As already mentioned, we assume that the set of patterns
{vrn ; n = 1� 2� . . . � N r − 4; r = 1� 2� . . . � 100}
represents the behavior of HR in healthy subjects and will be used
as a catalogue. One of the main motivations for this study is to
compare patterns obtained from the records before the arrhyth-
mias with patterns in the catalogue. Therefore, a primary objective
is to quantify the diversity of this set.
To this purpose, we focus on estimating how close were these
patterns to each other by computing the minimum distance from
each pattern (vrn ) to all others patterns (except for those in the same
rth record) in the catalogue:
ε(vrn ) = Min{D(vrn � vsm )}; s=
� r; 1 ≤ m ≤ N s − d � 1; 1≤s≤S
(2)
 Author's personal copy

E. Álvarez et al. / Medical Engineering � Physics 32 (2010) 1131–1136 1133

To establish a criterion for large values of ε(vrn ), we use the

Fig� 2� Percentage distribution P, of minimum distances εmin , between each pattern
in control group and its nearest neighbor. For d = 5 we obtain Nr − 4 vectors from each
record. Since the mean value over the 100 records is � Nr �≈4500 we get S ≈ 4.5 × 105
patterns in this group.
distribution of values of these distances.
The distribution of percentages of the S values of these minimum
distances corresponding to the case of d = 5 is shown in Fig. 2. As can
be seen, it is unlikely to find patterns whose distance to its near-
est neighbor is greater than 60 ms. In fact, this value corresponds
to 99.5th percentile and we used it as a threshold to differentiate
between large and small values of ε(vrn ).
There are a total number of S ≈ 4.5 × 105 patterns in this study
group; hence, there exist approximately 2250 vectors which are
singular patterns representing unlikely behavior of HR in healthy
subjects.
We end this section with a comment on the value of d. It is
clear that results must depend quantitatively on the value of this
parameter. For that reason, values in the range 1 ≤ d ≤ 10 were also
investigated. However, since we did not observe significant quali-
tative changes when 5 ≤ d ≤ 10, hereafter only results for d = 5 are
reported.
3� Atypical patterns before Paroxysmal Ventricular
Tachycardia

The histogram in Fig. 2 provides a criterion to establish whether

where a given pattern vx = (�x1 � �x2 � . . . � �x5 ) from xth record before PVT,
behaves as those of healthy subjects or not. According to Eq. (3),
�
100 � d−1
2 the minimum distance between the new pattern and the database
� ��
S= (N q − d � 1); r s
D(vn � vm ) = � ((RR)rn�i − (RR)sm�i )
is computed by means of:
q=1 i=1
(3) ε(vx ) = Min{D(vx � vsm )}; 1 ≤ m ≤ N s − d � 1; 1≤s≤S (4)
where
Clearly, a small value of ε(vrn ) means that there exists at least
�
� i=5
one approximate pattern to vrn , while a large value indicates that vrn �� 2
x s
D(v � vm ) = � (�xi − (RR)sm�i−1 ) (5)
is an isolated pattern, representing a singular and unlikely behavior
of HR. i=1

Fig� 3� Spatial distribution of projections of atypical patterns. From a record with 500 atypical patterns before PVT, we show projections on 3D for different sets of components:
1, 2 and 3 (a); 2, 3 and 4 (b); 3, 4 and 5 (c) and 2, 4 and 5 (d). In all cases it is clearly observed a clustered distribution.
 Author's personal copy

1134 E. Álvarez et al. / Medical Engineering � Physics 32 (2010) 1131–1136

Table 1
Classification of records before PVT and control group for a threshold U = 60.

Results for U = 60 Before PVT Control group

Positive (≥U) 70 (true positives) 12 (false positives)

Negative (≤U) 18 (false negatives) 88 (true negatives)

Table 2
Classification of records before PVT and Physionet control group for a threshold
U = 60.

Results for U = 60 Before PVT Physionet group

Positive (≥U) 70 (true positives) 34 (false positives)

Negative (<U) 18 (false negatives) 181(true negatives)

and then, if ε(vx ) is greater than 60 ms, the pattern is considered

atypical.
We observed a greater number of atypical patterns before
arrhythmias when compared with records of healthy subjects. Sep-
arately on each group we determined the number of records having
more(less) atypical patterns than a specified cutoff value U. Taking
U = 60 (that is, an average of one atypical pattern per minute) we
classified records having more(less) atypical patterns than U as pos-
itives(negatives), and we get the values as seen in Table 1, leading
to 79.5% of sensibility and 88% of specificity.
Similar results were obtained when the evaluation was per-
formed on a test set different from the training set used to compute
the threshold minimum distance. For instance, using a test set
from the Physionet database (Normal Sinus Rhythm RR Interval
Database (nsr2db), record: nsr2db/nsr001–nsr054, [21]) (215 ran-
domly selected hours from 54 healthy control subjects), we obtain
the values as seen in Table 2, leading to 79.5% of sensibility and 84%
of specificity.
It should also be noted that we did not find a relationship
between atypical patterns and ectopic beats. Indeed, we found
records without ectopic beats and a significant number of atypical
patterns (≈43%), as well as records with far fewer atypical patterns
that ectopic beats.
Next, we investigated the properties of these anomalous pat-
terns. Specifically, whether they have some particular profile or, on
the contrary, it is unlikely to find any atypical pattern that occurs
repeatedly within the same record. In the latter case, a nearly uni-
form spatial distribution of these vectors should be expected in the
region of space where they lie. Furthermore, spatial clustered dis-
tributions (i.e. with groups of nearly equal atypical patterns) might
be indicative that underlying mechanisms governing the onset of
arrhythmias are difficult to assimilate as uncorrelated random pro-
cesses.
Fig� 5� Characterizing spatial distribution of atypical patterns. Setting ı = 30 ms, each
axis is segmented in 10 intervals generating 105 boxes. Counting the number nxi of
data points within the ith box, we compute the proportion qxi = nxi /Ax , where A is
the total number of atypical patterns in the xth record.
For each record x we have extracted all its atypical patterns:
ax1 � ax2 � . . . � axk � . . . � axAx .
As illustrated in Fig. 3, these vectors are not uniformly dis-
tributed. In fact, in all records analyzed, a clustered distribution
of patterns with a relatively small number of clusters containing
many points was observed.
In order to quantify this clustering, the region containing atyp-
ical patterns was partitioned into small hypercubes of edge length
ı (Fig. 4), and we counted the number of points within each box.
Results of this counting procedure on records before PVT are
shown in Fig. 5. As can be observed, a relatively small number of
boxes contain the majority of points, corresponding to clusters in
the spatial distribution of atypical patterns. Even though results of
only one example are shown, the behavior is common to all records
and differences are mainly on the number of clusters observed
(which is always less than 10).
An interesting feature is observed when following the evolution
of patterns (see Fig. 6). In any short interval of time at the beginning
of the hour before the arrhythmia, the dynamics of atypical patterns
is confined within the clusters. Nevertheless, if the time interval is
at the end of the hour, there is an increasing number of points that
lie outside of the clusters. In other words, the evolution of atypical
patterns take place inside the clusters until it becomes random, and
patterns start to appear outside of the clusters in a noisy manner
shortly before the onset of the arrhythmia.

Fig� 4� Counting data points within boxes of edge length ı. When distributions are spatially clustered, there are relatively few boxes with many points inside. If the shape of
patterns is random in nature, it is unlikely to find repeated points and it should be expected a nearly uniform number of points within the boxes.
 Author's personal copy

E. Álvarez et al. / Medical Engineering � Physics 32 (2010) 1131–1136 1135

Fig� 6� Dynamical evolution of atypical patterns. Projection of 2500 atypical patterns from PVT record (a). The first 250 points (b) and the 250 at the middle (c) are grouped
in clusters while a significant proportion of the last 250 patterns (d) reside outside of the clusters.

Finally, besides calculating the number of atypical pat-
terns, we compared the distances between the catalogue and
the patterns at the beginning and the end of records before
arrhythmias.
As illustrated in Fig. 7, for each record x we computed the
number of atypical vectors nxit as well as the mean distance
�εxi �t on a window of size t at the beginning, and also the num-
ber of atypical vectors nxft and the mean distance �εxf � on a
t

Fig� 7� Chronological order of values of distances between catalogue and patterns. Four examples are displayed for a fixed size of windows (10 min). For each record we show
the number of atypical vectors na� as well as the mean distance �ε � on windows at the beginning and the end of the record. In all cases the number of atypical patterns and
the mean distance is greater close to the arrhythmia.
 Author's personal copy
1136 E. Álvarez et al. / Medical Engineering � Physics 32 (2010) 1131–1136
window of the same size at the end of the record, just before the
event.
For a mean size value of windows of 10 min, approximately 70%
of records satisfy:
nxft �εxf �
> 1; t
>1
nxit �εxi �t
meaning that differences between the catalogue and the patterns
are prone to increase close to the onset of arrhythmias.
These observed changes in the dynamic evolution of the patterns
near the arrhythmia could be useful in designing early warning
systems. Indeed, this is an issue on which we develop our ongoing
research.
4� Conclusions
In order to understand the underlying mechanisms of arrhyth-
mogenesis, the behavior of Heart Rate data from 1-h records
previous to Paroxysmal Ventricular Tachycardia was analyzed and
compared to those from healthy subjects.
Previous studies have reported statistically significant changes
in Heart Rate Variability parameters before arrhythmias. Oth-
ers, however, found no consistent variations before arrhythmias
of patients with moderate-severe cardiac disease [5–8,15–18].
Although we do not display results in detail, standard HRV param-
eters from the time and frequency domain were computed in order
to find the differences between PVT and control group. All of these
parameters including cycle length, standard deviation, root mean
square deviation, PNN50, low frequency, high frequency and the
ratio low/high frequencies were calculated on a number W of
windows of size T. For a range of values of W and T we found
that neither in the time nor in the frequency domain there are
significant changes in the parameters of the PVT group near the
arrhythmia.
This is the main reason why we consider patterns of Heart Rate.
There is a unique scalar quantity which is involved in the evalua-
tion of conventional indices: the values of RR intervals. Patterns are
defined as sequences of d consecutive values of RR intervals. Since
they are defined as vectors, they take into account not only values
of RR interval but also the chronological order in which they occur.
After constructing all possible patterns with five RR consecu-
tive intervals, comparisons were made by calculating the distance
between the patterns before arrhythmias and those of healthy
volunteers. Any pattern lying at a distance greater than a fixed
threshold (where we used 60 ms corresponding to 99.5th percentile
in control group) is named atypical.
Although the number of records is relatively small and should
be increased in subsequent work, there were two main results. On
one hand, all the healthy people have few atypical patterns when
compared with records before PVT. Drawing on Table 1, an episode
of PVT occurs 85.36% of times that the cutoff value is exceeded,
while no episode of PVT takes place 83% of times that the cutoff is
not exceeded.
On the other hand, these patterns follow a rather complex evo-
lution. Far from the onset of arrhythmia they can be grouped into a
small number of clusters which depend on the patient and, as the
onset of the arrhythmia approaches, the number of atypical vectors
per time unit falling outside the clusters increases. This change in
the nature of the dynamic evolution of patterns that occur near the
arrhythmia leads to an increasing of the number of atypical pat-
terns and of the average distance between atypical patterns and the
catalogue. This could be useful in designing early warning systems.
As already mentioned in Section 1, our method is comple-
mentary to other existing methodologies for risk stratification of
arrhythmias, which establish criteria to identify those patients who
have a high probability of being affected by an arrhythmia, but
without indicating when. Our efforts are aimed precisely at this
(6) point, that is, given a patient with a high risk of PVT, determining
how likely it is that an arrhythmic event will occur during the next
hour.
Acknowledgments
This work was partially supported by Consejo de Investigación
Cientfica y Humanstica, Universidad Central de Venezuela. Grant
PG 03-00-7189-2008.
Conflict of interest
None declared.
References
[1] Glass L. Understanding and preventing sudden cardiac death. In: Proc. 30th
Ann. Int. Conf. IEEE Eng. Med. Biol. Soc. 2008, clvii–clvii.
[2] Glass L, Lerma C. Risk stratification for arrhythmic sudden cardiac death. Heart
Rhythm 2006;3:1497–501.
[3] Huikuri HV, Perkiömäki J, Maestri R, Pinna G. Clinical impact of evaluation of
cardiovascular control by novel methods of heart rate dynamics. Philos Trans
R Soc A 2009;367:1223–38.
[4] Wilkoff BL, Williamson BD, Stern RS, Moore SL, Lu F, Lee SW, et al.
Strategic programming of detection and therapy parameters in implantable
cardioverter-defibrillators reduces shocks in primary prevention patients. J Am
Coll Cardiol 2008;52:541–50.
[5] Chesnokov YV, Holden AV, Zhang H. Distant prediction of paroxysmal atrial
fibrillation using hrv data analysis. Comput Cardiol 2007;34:455–8.
[6] Huikuri HV, Raatikainen MP, Moerch-Joergensen R, Hartikainen J, Virtanen V,
Boland J, et al. Cardiac arrhythmias and risk stratification after acute myocar-
dial infarction (carisma) study group. Prediction of fatal or near-fatal cardiac
arrhythmia events in patients with depressed left ventricular function after an
acute myocardial infarction. Eur Heart J 2009;30:689–98.
[7] Shin D, Yoo C, Yi S, Bae J, Kim Y, Park J, et al. Prediction of paroxismal atrial
fibrillation using non linear analysis of the r-r interval dynamics before the
spontaneous onset of atrial fibrillation. Circ J 2006;70:94–9.
[8] Thong T, McNames J, Aboy M, Goldstein B. Prediction of paroxysmal atrial fib-
rillation by analysis of atrial premature complexes. IEEE Trans Biomed Eng
2004;51:561–9.
[9] Alcaraz R, Rieta J. Sample entropy of the main atrial wave predicts spontaneous
termination of paroxysmal atrial fibrillation. Med Eng Phys 2009;31:917–22.
[10] Sun R, Wang Y. Predicting termination of atrial fibrillation based on the struc-
ture and quantification of the recurrence plot. Med Eng Phys 2008;30:1105–11.
[11] Watson J, Addison P, Uchaipichat N, Shah A, Grubb N. Predicting termination
of atrial fibrillation based on the structure and quantification of the recurrence
plot. Comput Biol Med 2007;37:517–23.
[12] Fischer R. Entrainment of the natural pacemakers of the heart precedes atrial
fibrillation. Comput Biol Med 2006;36:1204–15.
[13] Alcaraz R, Rieta J. Non-invasive organization variation assessment in the onset
and termination of paroxysmal atrial fibrillation. Comput Methods Programs
Biomed 2009;93:148–54.
[14] Zemlin CW, Mitrea BG, Pertsov AM. Spontaneous onset of atrial fibrillation.
Physica D 2009;238:969–75.
[15] Baumert M, Wessel N, Schirdewan A, Voss A, Abbott D. Scaling characteristics of
heart rate time series before the onset of ventricular tachycardia. Ann Biomed
Eng 2007;35:201–7.
[16] Burri H, Chevalier P, Arzi M, Rubel P, Kirkorian G, Touboul P. Wavelet trans-
form for analysis of heart rate variability preceding ventricular arrhythmias in
patients with ischemic heart disease. Int J Cardiol 2006;109:101–7.
[17] Chen S. A wavelet-based heart rate variability analysis for the study of nonsus-
tained ventricular tachycardia. IEEE Trans Biomed Eng 2002;49:736–42.
[18] Reed MJ, Robertson C, Addison P. Heart rate variability measurements and the
prediction of ventricular arrhythmias. QJM 2005;98:87–95.
[19] Arzeno NM, Deng Z, Poon C. Analysis of first-derivative based qrs detection
algorithms. IEEE Trans Biomed Eng 2008;55:478–84.
[20] Kohler BU, Hennig CR. Orglmeister, the principles of software qrs detection.
IEEE Eng Med Biol Mag 2002;21:42–57.
[21] Goldberger AL, Amaral LAN, Glass L, Hausdorff JM, Ivanov PC, Mark RG,
et al. PhysioBank, PhysioToolkit, and PhysioNet: components of a new
research resource for complex physiologic signals. Circulation 2000;101(June
(23)):e215–20, http://circ.ahajournals.org/cgi/content/full/101/23/e215.