FREQUENCY DOMAIN ANALYSIS AS RISK PREDICTOR

OF SUDDEN CARDIAC DEATH FROM LONG-TIME

ECG RECORDINGS

Diego Garc

ıa, C

esar S

anchez and Ra

ul Alcaraz

Innovation in Bioengineering Research Group, University of Castilla-La Mancha, Cuenca, Spain

Keywords:

Sudden cardiac death, Signal processing, Electrocardiogram, Frequency analysis.

Abstract:

Sudden Cardiac Death (SCD) is a disease that may not only affect patients with cardiovascular pathologies,

but also to apparently healthy patients. Thereby, identiﬁcation of patients with a high potential of suffering

SCD is crucial for their treatment with adequate therapies. To this respect, in the present work, different

signal processing tools were applied to surface electrocardiographic (ECG) recordings to develop markers

which can clearly differentiate between subjects without cardiovascular pathologies and patients who died of

SCD. Precisely, the proposed indexes were the Spectral Concentration (SC) around the main frequency peak,

which reached a sensitivity of 100.00% and a speciﬁcity of 88.89%, and the Mean Frequency Distance (MFD)

between the ﬁrst spectral peaks, which provided a sensitivity of 95.00% and a speciﬁcity of 100.00%.

1 INTRODUCTION

Sudden cardiac death (SCD) is nowadays a surprising

episode that causes death to 350,000 people per year

only in the USA (Al-Khatib et al., 2007). Accord-

ing to recent statistics, it is one of the main causes of

death for developed countries in apparently healthy

population (Chugh et al., 2000). Moreover, 50% of

deaths in patients with cardiovascular pathologies suf-

fered SCD (Al-Khatib et al., 2007). Only 5% of

patients who experienced SCD survived (Al-Khatib

et al., 2007), but many of them could have been saved

by cardioversion, especially those for which SCD was

the ﬁrst symptom.

Hence, risk stratiﬁcation is a determining factor to

reduce sudden cardiac mortality. To this respect, sev-

eral medical diagnosis techniques are currently used.

Thus, a low left ventricular ejection fraction (≤ 35 %)

is the gold standard, and subjects with NYHA class

II and III symptoms are at higher risk for SCD (Al-

Khatib et al., 2007). However, these methods have a

low predictive capacity (Arya et al., 2006). Thereby,

during the last years, many new lines of investigation

have been developed to ﬁnd effective markers which

are able to predict SCD with anticipation. Most of the

recently proposed predictors are based on the time do-

main analysis, such as QRS duration, QT dispersion,

ST abnormalities, T-wave alternans, late potentials,

or heart rate variability complexity (Al-Khatib et al.,

2007; Engel et al., 2004). Nevertheless, these indica-

tors have also shown a low diagnostic accuracy (Durin

et al., 2008). Overall, in this work, two promising in-

dexes, developed from the frequency domain analy-

sis of ECG signals, are proposed to detect subjects at

high risk.

2 MATERIALS

For the study, two databases available from Phys-

ioBank (Goldberger et al., 2000) were used: the MIT-

BIH Normal Sinus Rhythm (NSRDB) and Sudden

Cardiac Death Holter (SCDHDB). The ﬁrst set in-

cludes ECG recordings, with a length between 20 and

24 hours, belonging to ﬁve men aged 26 to 45 and

thirteen women aged 20 to 50. They are referred to

the Arrhythmia Laboratory at Beth Israel Deaconess

Medical Center, and were found not to have signiﬁ-

cant arrhythmias. This group of 18 ECG recordings

of normal sinus rhythm was considered as reference.

The second database contains 23 Holter signals asso-

ciated to patients who suffered SCD during the 24-h

monitoring. They were mainly obtained in the 1980’s

in Boston area hospitals. In the analysis, 20 record-

ings were only considered, given that paced patients

were discarded.

424

García D., Sánchez C. and Alcaraz R. (2010).

FREQUENCY DOMAIN ANALYSIS AS RISK PREDICTOR OF SUDDEN CARDIAC DEATH FROM LONG-TIME ECG RECORDINGS.

In Proceedings of the Third International Conference on Bio-inspired Systems and Signal Processing, pages 424-427

DOI: 10.5220/0002742604240427

 SciTePress

3 METHODOLOGY

To predict SCD, the 60 minutes preceding ventricu-

lar ﬁbrillation (VF) were taken from each SCDHDB

recording. For the NSRDB signals, 60 minutes start-

ing at an random instant were selected. In order to

unify the sampling rates of the considered segments,

the ﬁrst lead was resampled at 1 kHz. Addition-

ally, to improve later analysis, interferences present

in the ECG were removed with four ﬁlters (S

ornmo

and Laguna, 2005): an 8th-order Chebyshev low-pass

ﬁlter ( f

= 100Hz), a 3rd-order Butterworth high-

pass ﬁlter ( f

= 0.50Hz), a 2nd-order IIR notch ﬁl-

ter ( f

= 60Hz), and an envelope detector for baseline

wander subtraction.

The power spectral density (PSD) of each chosen

segment was computed over 4096 samples-length in-

tervals, obtained with a Hamming window and over-

lapped 2048 samples, making use of an 8192-points

Fast Fourier Transform (FFT). Finally, the Spectral

Concentration (SC) around the main frequency peak

and the mean frequency distance (MFD) between the

N ﬁrst were computed for each interval, thus obtain-

ing two numerical series.

3.1 SC Around the Main Frequency

Peak

The Spectral Concentration, SC[x(t)], around the

main frequency peak, f

, has been previously used

as a performance indicator of the atrial activity ex-

traction in atrial ﬁbrillation and tachyarrhythmias

episodes (S

anchez et al., 2004; Castells et al., 2005).

It has been deﬁned as:

SC[x(t)] =

1,17· f

∑

f =0,82·f

( f )

∑

f =0

( f )

(1)

where f

is the sampling rate of the analyzed segment,

x(t) and x

( f ) is its PSD: x

( f ) = ℜ



DFT {x(t)}



This index relates present energy on the peak band to

the rest of the distribution of interest. In this sense,

an ideal sinus signal has a SC of 1, because all spec-

tral content is on its own oscillation frequency. Simi-

larly, ECG signal with a ventricular ﬂutter presents a

SC value very high, because this arrhythmic episode

is highly regular (Castells et al., 2005). On the con-

trary, low SC values means that only an small part of

the PSD is concentrated around main frequency peak,

which suggests the existence of other notable peaks in

the rest of the spectrum.

3.2 MFD between the First Spectral

Peaks

The MFD[x(t),N] is deﬁned as the mean frequency

distance between the N ﬁrst peaks of the PSD. Its

computation requires a smooth spectrum to determine

with high accuracy the frequency peaks. Because of

the PSD provided by Fast Fourier Transform (FFT)

shows transitory peaks that may distort the index out-

come, a smoothing spectrum technique consisting in

a low-pass ﬁlter with a 40 samples-length Hamming

window was applied (Proakis and Manolakis, 2007).

This parameter can be used to estimate the PSD dis-

persion. Thus, a low value implies close frequency

peaks and, therefore, an concentrated spectrum. On

the contrary, a high value indicates a high dispersion

and a PSD with considerable empty gaps between fre-

quency peaks.

4 RESULTS

Table 1 shows the obtained SC and MFD values for

all the recordings. To associate every ECG with an

only value, the mean for the time series provided by

each parameter was computed. The differences be-

tween the values obtained for the two groups, see

Fig. 1, were statistically signiﬁcant, given that the sig-

niﬁcance level, obtained by a t Student test, was p <

0.0001 for both markers. Receiver Operating Charac-

teristics (ROC) curve was used to obtain the discrim-

ination threshold between healthy patients and those

who died by SCD (Lasko et al., 2005). In addition,

the predictive ability of both indicators was also ob-

tained with this curve. Thus, sensitivity was consid-

ered as the number of healthy patients correctly clas-

siﬁed, whereas speciﬁcity represented the percentage

of patients who died by SCD correctly discerned. Fig

2 shows the ROC curves for the two makers, and Ta-

ble 2 presents their values obtained of sensitivity and

speciﬁcity.

As can be appreciated, the SC revealed only two

falses positives (signals #19140 and #19093) with an

area under ROC curve (AUC) of 0.9694 and an ac-

curacy of 94.74% (36 out of 38). WIth regard to the

MFD, only one false negative (signal #39), an AUC

of 0.9944 and an accuracy of 97.37% (37 out of 38)

were provided. Since the failures obtained with both

tests were associated to different signals, a combin-

ing discriminator considering a positive result with

SC > 16.25% and MFD > 3.75Hz, would have an ac-

curacy of 100% (38 out of 38).

FREQUENCY DOMAIN ANALYSIS AS RISK PREDICTOR OF SUDDEN CARDIAC DEATH FROM LONG-TIME

ECG RECORDINGS

425

Table 1: SC and MFD results for all the analyzed record-

ings.

NSRDB SCDHDB

Signal SC MFD Signal SC MFD

# (%) (Hz) # (%) (Hz)

16265 11.9 2.704 30 26.1 5.848

16272 10.1 2.289 31 27.0 5.070

16273 12.2 3.404 32 49.4 5.747

16420 14.0 1.719 33 20.5 4.798

16483 11.5 2.423 34 16.3 5.960

16539 10.5 1.988 35 27.0 4.327

16773 12.7 3.070 36 19.1 5.514

16786 12.9 3.705 37 25.1 4.453

16795 14.9 2.917 38 67.7 4.523

17052 11.3 1.656 39 25.8 3.788

17453 16.1 3.340 41 34.5 5.154

18177 12.0 2.428 43 22.3 3.777

18184 12.1 1.729 44 16.3 5.022

19088 12.7 2.722 45 16.9 4.619

19090 10.9 2.228 46 16.4 3.952

19093 18.4 2.068 47 37.8 4.294

19140 22.2 3.661 48 24.0 4.428

19830 16.2 1.901 50 24.5 4.967

51 33.8 4.314

52 20.9 4.214

Mean 13.5 2.553 Mean 27.5 4.724

Std 3.1 0.674 Std 8.2 0.685

(a) SC (b) MFD

Figure 1: SC and MFD for NSRDB and SCDHDB (‘box-

and-whiskers’ plot).

4.1 Numerical Evolution of Markers

Given that results showed that SCD could be suc-

cessfully predicted with the SC and MFD, both pa-

rameters were computed for a longer time interval.

Thus, the 300 minutes preceding VF from 13 SCD-

HDB were selected. In addition, for all the NSRDB

signals, 300 minutes starting at a random instant were

also chosen. For this case, the time course provided

Table 2: Area Under ROC Curve (AUC), threshold, sensi-

tivity and speciﬁcity for SC and MFD tests.

AUC Threshold Sensitivity Speciﬁcity

SC 0.9694 16.25 % 100.00% 88.89%

MFD 0.9944 3.750 Hz 95.00% 100.00%

Figure 2: ROC Curves.

Figure 3: SC numerical evolution.

Figure 4: MFD numerical evolution.

BIOSIGNALS 2010 - International Conference on Bio-inspired Systems and Signal Processing

426

by both markers is displayed in Figs. 3 and 4.

5 DISCUSSION AND

CONCLUSIONS

In the present work spectral processing has been used

to obtain two indexes that reveal statistically signiﬁ-

cant differences (p < 0.0001) between patients who

suffered SCD and a sample of completely healthy

subjects.

Both parameters showed values substantially

higher for patients who suffered SCD than for healthy

subjects. This fact suggests that the ECG of healthy

subjects is characterized by a higher harmonic con-

tent. Consequently, it could be considered that fre-

quency peaks of relative high amplitude in the ECG

of healthy patients are disappeared in subjects who

suffered SCD, such as Fig. 5 shows.

Figure 5: Typical PSD of an ECG for the (a) NSRDB and

(b) SCDHDB.

On the other hand, the time course of the markers

showed very constant values during the last 5 previous

hours preceding the death of the patient. Therefore,

SCD could be predicted with an anticipation above 5

hours, which suggests the problem that causes SCD,

it can be congenital or acquired during subject’s life.

Finally, given that the analyzed database are lim-

ited, the results should be considered with caution.

Nevertheless, the work suggests that SC and MFD can

initiate new lines of research as non-invasive predic-

tors of SCD. In this sense, a wider data set allowing

a more rigorous statistical analysis should be required

in order to provide conﬁdence in the robustness of the

proposed parameters.

ACKNOWLEDGEMENTS

This work was supported by the projects PII1C09-

0036-3237 and PII2C09-0224-5983 from Junta de

Comunidades de Castilla-La Mancha.

The authors would like to acknowledge the invalu-

able helpful support received from Dr. J.L. Bardaj

ı,

Dr. M.L. L

opez, Dr. F. Madero, and Eng. M.E.

Garc

ıa.

REFERENCES

Al-Khatib, S. M., Sanders, G. D., Bigger, J. T., Buxton,

A. E., and et al (2007). Preventing tomorrow’s sud-

den cardiac death today: part i: Current data on risk

stratiﬁcation for sudden cardiac death. Am Heart J,

153(6):941–950.

Arya, A., Haghjoo, M., and Sadr-Ameli, M. A. (2006). Risk

stratiﬁcation for arrhythmic death after myocardial in-

farction: current perspective and future direction. Int

J Cardiol, 108(2):155–164.

Castells, F., Rieta, J. J., Millet, J., Zarzoso, V., and Asso-

ciate (2005). Spatiotemporal blind source separation

approach to atrial activity estimation in atrial tach-

yarrhythmias. IEEE Trans Biomed Eng, 52(2):258–

267.

Chugh, S. S., Kelly, K. L., and Titus, J. L. (2000). Sudden

cardiac death with apparently normal heart. Circula-

tion, 102(6):649–654.

Durin, O., Pedrinazzi, C., Donato, G., Pizzi, R., and Inama,

G. (2008). Usefulness of nonlinear analysis of ecg

signals for prediction of inducibility of sustained ven-

tricular tachycardia by programmed ventricular stim-

ulation in patients with complex spontaneous ventric-

ular arrhythmias. Ann Noninvasive Electrocardiol,

13(3):219–227.

Engel, G., Beckerman, J. G., Froelicher, V. F., Yamazaki,

T., and et al (2004). Electrocardiographic arrhythmia

risk testing. Curr Probl Cardiol, 29(7):365–432.

Goldberger, A. L., Amaral, L. A., Glass, L., Hausdorff,

J. M., Ivanov, P. C., and et al (2000). Physiobank,

physiotoolkit, and physionet: Components of a new

research resource for complex physiologic signals.

Circulation, 101(23):E215–E220.

Lasko, T. A., Bhagwat, J. G., Zou, K. H., and Ohno-

Machado, L. (2005). The use of receiver operat-

ing characteristic curves in biomedical informatics. J

Biomed Inform, 38(5):404–415.

Proakis, J. G. and Manolakis, D. K. (2007). Digital Signal

Processing. Principles, Algorithms and Applications.

Pretence Hall.

anchez, C., Rieta, J. J., Castells, F., Alcaraz, R., and Mil-

let, J. (2004). Wavelet domain blind signal separation

to analyze supraventricular arrhythmias from Holter

registers. Conf Proc Independent Component Analy-

sis and Blind Signal Separation, 5:1111–1117.

ornmo, L. and Laguna, P. (2005). Bioelectrical Signal Pro-

cessing in Cardiac and Neurological Applications. El-

sevier Academic Press.

FREQUENCY DOMAIN ANALYSIS AS RISK PREDICTOR OF SUDDEN CARDIAC DEATH FROM LONG-TIME

ECG RECORDINGS

427