Classification System for Fetal Heart Rate Variability Measures … · 2013. 11. 27. · University...
Transcript of Classification System for Fetal Heart Rate Variability Measures … · 2013. 11. 27. · University...
Classification System for Fetal Heart Rate
Variability Measures Based on
Cardiotocographies
João A. L. Marques University Lusiada of Angola/Department of Informatics, Lobito, Angola
University of Leicester/Department of Engineering, Leicester, United Kingdom
Email: [email protected]
Paulo C. Cortez and João P. V. Madeiro Federal University of Ceará/Department of Engineering of Teleinformatics, Fortaleza, Brazil
Email: {cortez, joaopaulo}@deti.ufc.br
Fernando S. Schlindwein University of Leicester/Department of Engineering, Leicester, United Kingdom
Email: [email protected]
Abstract—The Fetal Heart Rate interpretation based on
Cardiotocographies (CTG) is the most common practice of
obstetrician medical staffs. Computerized CTG Systems are
used with the aim to reduce subjective aspects of these
diagnostics. The Fetal Heart Rate Variability (FHRV)
analysis using the CTG signal is an unusual approach. This
work proposes a FHRV analysis based on the evaluation of
time domain parameters (statistic measures); frequency
domain parameters; and the short and long term variability
obtained from the Poincaré plot. A normal distribution is
presumed for each parameter and a normality criterion is
proposed. Specific and overall classifications are proposed
to help improve the fetal conditions interpretation,
expanding the conventional FHR analysis.
Index Terms—fetal heart rate variability (FHRV),
cardiotocography (CTG), diagnostic
I. INTRODUCTION
The cardiologic and autonomic nervous systems (ANS)
continuously look for a dynamic balance where the
parasympathetic and sympathetic systems act as opposite
forces influencing the heart rhythm modulation. The first
one increases the heart rate and decreases the variability
while the second system does the opposite action [1].
The Fetal Heart Rate Variability (FHRV) can be
obtained by the Cardiotocography (CTG), as it is
considered a gold standard exam for the detection of fetal
heart rate (FHR). The Doppler sensor has similar
accuracy when compared with the abdominal ECG for
the fetal heart beat detection [2].
The CTG records continuously and simultaneously the
FHR and the uterine tonus (for uterine contractions
monitoring). Fetal movements can also be recorded
Manuscript received June 3, 2013; revised August 25, 2013.
manually by the mother. These monitoring allow the
detection of a large set of diseases or changes in the fetal
health status [3].
Usually, the CTG exam is done in risky pregnancies
because fetal distress can be earlier detected. Depending
on the situation, the exam is applied before labour, period
of time named as antepartum, and also during labour, the
intrapartum period [4].
Previous works are using the FHVR analysis acquiring
the fetal ECG Signal. Lebrun (2003) states that the FHRV
analysis during the last trimester can provide important
clinical information after birth [5]. A fetal development
indice based on time and frequency parameters is
suggested based on the FHR decrease and variability
increase during the pregnancy.
Sibony et al. (1994) present that the FHRV can be used
to detect the fetal status also during labour [6]. They
propose the identification of new frequency intervals and
two evaluation criteria based on the FHRV spectral
analysis. Other works consider the frequency domain
parameters as part of an overall comparison with other
monitored systems to determine fetal status [7].
The time domain HRV analysis considers geometric
and statistic approaches [8]. The geometric metrics are
based on the histogram of the set of normal intervals
between QRS complexes. In statistical analysis there are
several metrics divided in three groups. Each metric is
defined in Table I. The first one evaluates the heart rate
behaviour as a whole, i.e., considers the whole set of
samples for its calculations. The metrics are the SDNN
and the SDANN. The RMSSD measure considers the
interval between heart beats and belongs to the second
group and reflects the high frequency characteristics of
the signal. Finally, the long term variability (LTV) and
short term variability (STV) are obtained from the
Poincaré plot.
2013 Engineering and Technology Publishing
Journal of Life Sciences and Technologies Vol. 1, No. 3, September 2013
184doi: 10.12720/jolst.1.3.184-189
The parasympathetic stimulation results in a fast and
short term answer in the heart beats, affecting
immediately the interval between them. It can be
evaluated when considering parameters such as RMSSD.
The sympathetic stimulation is slower with a latency
period that can vary from 5 to 20 seconds [1]. Parameters
considering many RR intervals, such as SDNN and
SDANN are used to evaluate both systems as a whole.
TABLE I. HRV MEASURES CONSIDERED AS CLASSIFICATION
CRITERIA
Measure Unit Description
SDNN ms Standard deviation considering
all NN intervals
SDANN ms Mean of all standard deviations of 5-minutes
segments of NN.
RMSSD ms The square root of the mean squared difference of
successive NNs
LF ms2 Low frequency power (0,04 – 0,15 Hz)
HF ms2 High frequency Power
(0,15 – 0,40 Hz)
LF/HF -- LF and HF high frequency ratio
The frequency domain parameters usually considered
in heart rate variability analysis are divided into
frequency intervals, such as Ultra Low Frequency (ULF),
Very Low Frequency (VLF), Low Frequency (LF) and
High Frequency (HF) [1]. In this work we consider only
the LF and HF intervals. The HF component corresponds
to changes in the heart rate related with the respiratory
cycles, which are tipically managed by the
parasympathetic system. On the other hand, the LF
component is influenced by both systems.
Actually, the HRV in time and frequency domain are
different expressions of the same phenomenon, some
correlation among those parameters can be demonstrated.
The SDNN parameter for example is related to the total
power of the spectral analysis. The time domain RMSSD
is correlated with the high frequency component in the
frequency domain since it considers the difference
between two RR adjacent intervals, quantifying fast
changes of the heart rate. Another example is the
correlation between the SDANN and the ULF frequency
band.
This work presents a FHRV analysis based on the
signal obtained by the CTG examination. Classification
criteria for the FHRV parameters and for the examination
as a whole are proposed.
II. MATERIALS AND METHODS
A. Database and Development Environment
The Matlab software version 7.6.0.324 R2008a is used
as the development environment [9]. The whole set of
HRV parameters (time and frequency domain) and the
STV and LTV based on Poincaré plot were calculated
using the system proposed by Madeiro [10].
The results were obtained from one previously
identified database from the Trium Analysis Online
GmBH, in Munich, Germany. The database is identified
as CTG-A, and has 80 examinations in antepartum period
of time, i.e., before labour, with gestational age varying
from the 28th
to the 34th
week.
From these, 58 examinations are classified as control
(normal fetal and low level of suspicious status) and 22 as
study (high level of suspicious or pathological). There are
no uterine contractions and the occurrence of FHR
accelerations may indicate normality.
B. FHRV Analysis
A block diagram with all the steps to perform the FHR
variability analysis is presented in Fig. 1. After the time
and frequency domain parameters are determined, the
long and short term variability can be obtained from the
Poincaré plot.
Figure 1. The block diagram of FHRV analysis
After calculating several different parameters, this
work considers the subset defined in Table I. The
frequency ranges considered in this work are also
presented. The LF and HF are expressed in normalized
units. The LF/HF ratio is considered in all the results
because it shows the balance between the sympathetic
and parasympathetic systems.
C. Classification Criteria
There are no previously determined normality criteria
for the FHRV. This work presents a classifier based on a
set of criteria.
Considering a normal statistical distribution for each of
the chosen parameters (Pi), the following criteria are
considered in this analysis:
Normality:
µPi - σPi ≥ Pi ≤ µPi + σPi
Suspicious:
µPi + σPi < Pi ≤ µPi + 2(σPi)
µPi - 2(σPi) ≤ Pi < µPi - σPi
Abnormality:
Pi > µPi + 2(σPi)
Pi < µPi - 2(σPi)
where µPi is the mean and σPi is the standard deviation for
each parameter Pi.
For the overall classification of the exam, four
different possibilities are considered:
If all parameters are classified as normal, then the
exam is considered as “FRHV Normality”.
2013 Engineering and Technology Publishing
Journal of Life Sciences and Technologies Vol. 1, No. 3, September 2013
185
If one parameter is suspicious then the exam is
labeled as “Attention”.
If two or more are suspicious, then the label is
“Suspicious”.
If any parameter is classified as abnormal, the
exam is classified as “FRHV Abnormality”.
It is important to notice that these criteria are not
comparable with the conventional classification and is
proposed to act as a complementary tool. The
conventional analysis can find pathologies where the
FHR does not and vice versa.
III. RESULTS AND DISCUSSION
In this section, the results for all of the parameters is
presented, such as for the time domain or the frequency
domain. The mean µ, the standard deviation σ, the
maximum value Max, the minimum value Min and the
Pearson variability are presented in Table II and Table III.
TABLE II. HRV MEASURES CONSIDERED AS CLASSIFICATION
CRITERIA
Measure SDNN SDANN RMSSD
µ 31.80 22.13 3.36
σ 10.81 8.67 0.96
Max 81.18 53.61 8.25
Min 15.07 5.05 1.93
Pearson
Variability 34.01 39.20 28.71
TABLE III. HRV MEASURES CONSIDERED AS CLASSIFICATION
CRITERIA
Measure SDNN SDANN RMSSD
µ 31.80 22.13 3.36
σ 10.81 8.67 0.96
Max 81.18 53.61 8.25
Min 15.07 5.05 1.93
Pearson
Variability 34.01 39.20 28.71
The Pearson Variability indicates that in this database
there are significant variations in almost the whole set of
parameters, especially when analysing the time domain
parameters. This indice is also high for the ratio LF/HF.
For a better comprehension of how the parameters are
related to each other the correlation coefficient ρ are also
calculated. For the long and short term variability
parameters, STV and LTV, is found ρ = 0.5739, showing
a strong correlation between them. For the SDNN and
SDANN parameters there is a stronger correlation, with ρ
= 0.6984. For the RMSSD parameter, which is by
definition, related with high frequency components, there
is a correlation with the HF parameter, ρ = 0.4706.
After these preliminary results, the detailed
classification process for each of the parameters is then
performed, assuming that all of them follow a normal
distribution.
In Fig. 2, the LF/HF values are plotted classified as
normal, suspicious and abnormal. The results according
to the normality classification contain the most part of the
exams, while only three were considered abnormal.
Figure 2. LF/HF classification
Figure 3. Varaiability classifications: (a) LTV and (b) STV
The scatter plot for the variability parameters, LTV
and STV, are presented in Fig. 3 (a) and Fig. 3 (b). There
are two significant outliers in both plots. These exams
must be carefully analysed as they may indicate fetal
distress.
Finally, the time domain statistics are presented in Fig.
4 (a), (b) and (c). As also shown in other previous
graphics, there are only a few abnormal and a significant
number of suspicious classification. Besides, there are not
abnormal exams under the lower suspicious values.
After this classification based on each parameter, the
overall analysis of each examination is performed. There
are four different outputs and the “Suspicious” was the
most common classification with 35% of the occurrences
while 21% were classified with the “Attention” label.
This means that 56% of the whole examinations set had
at least one FRHV parameter out of the normality
classification criteria. There were 30% of “Normal” and
14% of “Abnormal” outputs. These results are presented
in Fig. 5.
A group of four exams previously classified as normal
are presented in Table IV. All of them are also classified
2013 Engineering and Technology Publishing
Journal of Life Sciences and Technologies Vol. 1, No. 3, September 2013
186
as normal for the proposed classification using the FHRV
parameters. For these exams, there is a match between the
visual analysis and the FHRV analysis, considering the
normality classification proposed in this work. In a
general point of view, there are 17 exams among 58 with
this match of classification. Also, eight exams previously
classified as normal received the “Attention” label.
Figure 4. Time domain classifications: (a) SDNN, (b) SDANN and (c) RMSSD
Figure 5. Overall classification
In Table V, a set of three exams previously classified
as control, i.e., pathological or suspicious. For these
exams, at least two parameters are in the suspicious
classification based on the proposed (µ ± σ) analysis. The
third exam, for example, has STV, RMSSD and LF/HF
out of the proposed normality interval.
All the three exams classified as abnormal according to
the FHRV analysis and also were classified as
pathological by the conventional CTG analysis:
ctg20011218_2348371; ctg20001213_0948395 and
ctg20000709_043356.
TABLE IV. STATISTICAL AND FREQUENCY DOMAIN PARAMETERS FOR
NORMAL EXAMINATIONS
Exam STV HF
(n.u.) LF/HF SDNN RMSSD
ctg20000304-
0409053 0.81 13.72 6.28 21.68 2.95
ctg20000209-
0834583 0.90 10.62 8.41 26.69 2.81
ctg20000202-
0408315 0.98 11.94 7.37 31.86 3.41
ctg20000228-
1258193 1.09 10.87 8.19 30.73 3.64
ctg20000329-
0541413 0.85 13.15 6.60 22.57 2.80
TABLE V. STATISTICAL AND FREQUENCY DOMAIN PARAMETERS FOR
SUSPICIOUS OR PATHOLOGICAL EXAMINATIONS
Exam STV HF
(n.u.) LF/HF SDNN RMSSD
ctg20000729-2151501 1.22 13.31 3.94 28.40 3.42
ctg20011204-0845235 0.78 9.51 9.50 21.72 2.18
ctg20000630_0916173 0.66 8.49 5.74 60.55 2.09
Nevertheless, the FHRV the results presents that those
parameters must not be used as the unique analysis of the
fetal state. Other exams show a divergence if you
compare the two classification system. For example, the
exams ctg20000521-1402455, ctg20000203-1942093 and
ctg20000518-2034363 belongs to the proposed normality
classification but were previously classified as suspicious
or pathological. In all these cases, if the FHRV was
considered alone, fetal health problems could not be
detected.
On the other hand, the FHRV analysis may expands
the conventional analysis. The ctg20010223-1429403
exam is previously classified as normal in the
conventional analysis. Although, this exam presents the
lowest short term variability, STV=0.61, and the lowest
RMSSD, 1.93. The LTV=10.69 and HF=8.49 are
considered low values. This may indicate low variability
and a very small contribution of the high frequency
components, related to the parasympathetic system.
According to this analysis, the exam could be classified
as suspicious or pathological.
Another example is the ctg20010626-2358115 exam,
also classified as normal before. It presents the highest
SDNN value, 81.18 and high values for the HF, 22.09
14%
35% 21%
30%
Overall Classification
Abnormal
Suspicious
Attention
Normal
2013 Engineering and Technology Publishing
Journal of Life Sciences and Technologies Vol. 1, No. 3, September 2013
187
and the RMSSD, 4.70. This may indicate a suspicious
fetal health status, with a strong influence of the
parasympathetic system over the desired balance.
IV. CONCLUSIONS
The interval between heart beats can be considered an
important parameter for the detection of the fetal health
status.
The FHRV analysis based on the CTG examination is
a viable approach for the obstetric practice, since the
visual analysis is very subjective. For computerized CTG
systems this analysis can be applied as a second level and
complimentary detector of fetal distress.
The results presented require that the FHRV analysis
must not be considered as the only monitored parameters.
The conventional analysis is strictly necessary.
Future works could consider other statistical indices,
such as percentiles or quartiles for the classification
criteria and also other statistical and geometrical
measures can be considered, to improve the proposed
analysis.
ACKNOWLEDGMENT
The authors thank the Trium Analysis Online GmBH
in Munich, CNPq (Brazil), Funcap/FINEP (Brazil) for
funding PAPPE Project and the Centre of Bioengineering
at University of Leicester, Leicester, UK.
REFERENCES
[1] Task Force of the European Society of Cardiology and the North
American Society of Pacing and Electrophysiology, Heart rate
variability-standards of measurement, physiological interpretation,
and clinical use, European Heart Jounal, vol. 17, pp. 354-381,
1996.
[2] M. Signorini, G. Magenes, S. Cerutti, and D. Arduini, “Linear and nonlinear parameters for the analysis of fetal heart rate signal from
cardiotocographic recordings,” IEEE Transactions on Biomedical Engineering, vol. 50, pp. 365-374, 2003.
[3] I. Ingemarsson, E. Ingemarsson, Spencer, A. D. John, Fetal Heart
Rate Monitoring–A Practical Guide, New York: Oxford Medical Publications, Oxford University Press, 1993.
[4] J. A. L. Marques, P. C. Cortez, and J. P. do V. Madeiro, “Detecção
de alterações da frequência cardíaca fetal e do tônus uterino materno em exames cardiotocográficos utilizando transformada de
hilbert. porto de galinhas,” Brazilian Conference in Biomedical
Engineering, 2008. [5] D. N. Lebrun, Analysis of Neonatal Heart Rate Variability and
Cardiac Orienting Responses, Dissertação (Mestrado)-University
of Florida, 2003. [6] O. Sibony, J. Fouillot, M. Benaoudia, A. Benhalla, J. Oury, C.
Sureau, and P. Blot, “Quantification of fetal heart rate variability
by spectral analysis of fetal well-being and fetal distress,” European Journal of Obstretics and Gynecology and Reproductive
Biology, vol. 54, pp. 103-108, 1994.
[7] M. Ferrario, M. Signorini, G. Magenes, and S. Cerutti, “Comparison of entropy-based regularity estimators: Application
to the fetal heart rate signal for the identification of fetal distress,”
IEEE Transactions on Biomedical Engineering, vol. 53, pp. 119-125, 2006.
[8] M. Malik and A. J. Camm, “Components of the heart rate
variability-What they really mean and what we really measure,” American Journal of Cardiology, vol. 72, pp. 821-822, 1993.
[9] Mathworks. Matlab. Nov. (2011). [Online]. Available:
http://www.mathworks.com
[10] J. P. do V. Madeiro, “Sistema automático para análise de
variabilidade da freqüência cardíaca,” Master s Thesis, Federal
University Federal of Ceará, 2007.
J. A. L. Marques was born in Fortaleza-CE, Brazil,
in 1973. Dr. Marques graduated in Electrical
Engineering at the Federal University of Ceará (UFC) in 1996. He then concluded his Master’s
studies at UFC in 2006 and his doctoral studies at
UFC in 2010 (with an internship at Trium Analysis Online GmBH, in Munich, Germany - 2007). He
concluded postdoctoral studies at the University of
Leicester, UK in 2012. He is currently with the Department of Informatics, University Lusíada of Angola, in Lobito, Angola, where he
heads the research group of Health Informatics and is the Director of the
Research, Studies and Post-Graduation Center. His current research is focused on biological time series analysis (such as heart, brain and
others) using digital signal processing techniques based on linear and
nonlinear approaches. He is also heading a research project: the “NeuroSapiens Project”, a joint research with the University of Cape
Town, South Africa, a biological signal monitoring and analysis system
for cognitive and affective neuroscience studies in Angola with post-civil war students and families.
P. C. Cortez was graduated in Electrical
Engineering at the Federal University of Ceará
(UFC) in 1982. He then concluded his Master’s and Doctoral studies at UFC in 1992 and 1996
respectivelly at Federal University of Paraiba –
Campina Grande. His is an Associate Professor Level III from the Department of Teleinformatics
Engineering at UFC. His research is focused on
Artificial Vision, primarily working with 2-D and 3-D contours poligonal modeling, pattern recognition, digital imaging
segmentation digital signal processing, biomedical images, computer-
aided intelligent systems for biomedical signal analysis, telemedicine applications and embedded systems.
J. P. do V. Madeiro was graduated in Electrical Engineering at the Federal University of Ceará
(UFC) in 2006. He then concluded his Master’s
and Doctoral studies at UFC in 2007 and 2013 also at the Federal University of Ceara – Department of
Teleinformatics Engineering. He also worked at
the University of Leicester during his Doctoral studies workins with electrograms during
persistent atrial fibrilation. He works at Ministério Público Federal and
his research is focused on digital signal processing, computer-aided diagnostic systems, automatic ECG parameter extraction, electrograms
and the application of nonlinear techniques for cardiologic signals.
Fernando S. Schlindwein was born in Porto Alegre, Brazil, in 1956. He graduated with a First
Class Honours degree as an Electronic Engineer in
1979 from the Federal University of Rio Grande do Sul, Brazil, with an extension degree in
Nuclear Engineering. After a short time in
industry (Aços Finos Piratini, a steel mill) he obtained an MSc in Biomedical Engineering from
the Coordination of Post-Graduation Programmes in Engineering of the
Federal University of Rio de Janeiro (COPPE/UFRJ), Brazil in 1982, a PhD in Biomedical Engineering from the Department of Surgery of the
University of Leicester, England in 1990, and a DSc in Biomedical
Engineering from the Federal University of Rio de Janeiro (UFRJ) in 1992. He was a Senior Lecturer associated with UFRJ from August
1980 until 1992, when he joined the Department of Engineering at the
University of Leicester where he is a Reader in Bioengineering. He did his military service at Colégio Militar of Porto Alegre where he was
First Cadet in Infantry. He has also been a Senior Lecturer of the
2013 Engineering and Technology Publishing
Journal of Life Sciences and Technologies Vol. 1, No. 3, September 2013
188
Department of Electronics of the Brazilian Navy Academy for Officers
in Rio de Janeiro, Brazil in the early 1980s. His current research
interests are real-time digital signal processing, with more intense
research activities in i) cardiac arrhythmias, especially atrial fibrillation;
ii) heart rate variability and automatic arrhythmia monitoring using the
ECG, and iii) microprocessor-, microcomputer- and Digital Signal
Processor-based systems.
2013 Engineering and Technology Publishing
Journal of Life Sciences and Technologies Vol. 1, No. 3, September 2013
189