Ventricular fibrillation detection by autocorrelation function peak analysis

Ventricular Fibrillation Detection by Autocorrelation Function Peak Analysis

S e r g i o G. G u i l l e n , E E , P h D , * M a r i a T. A r r e d o n d o , Gregorio

Martin,

MS, PhD,t

and Jose M. Fe~ero

EE, PhD,*

Corral, EE, PhD*

Abstract: The use of reliable automatic ventricular fibrillation (VF) recognition techniques is critical in performing external automatic defibrillation. The authors' objective was to develop a method to detect VF and life-threatening arrhythmias, based on direct and simple peak analysis of the autocorrelation function (ACF). This method may differentiate between fibrillating and nonfibrillating rhythms, and in the first case between "coarse" and "fine" VF. ECG records during ventricular tachycardia (VT) and VF were obtained from patients during cardiac surgery. Segments 4 sec long were selected from tapes and digitized at 200 Hz, then split into three groups (VT, VF regular waveform, and VF irregular waveform). The positive peak P(j) of the ACF was defined as the maximum value between two function zeros, and RPL(j) represents the relation between P(j) and twice their own standard error. Parameter TR(1) was defined as the relation between P(1) width and the time of occurrence. ACFs were computed for the entire sample; RPL(j), D(j) = RPL(j) - RPL(j + 1), and TR(1) were calculated for every record. The results indicate that: (A) If RPL(1) > 1.2 and (1.6 > RPL(2) > l) and (RPL(3) > 0.6 and D(1) > 0), then consider VT; (B) If (1.2 > RPL(1) --> 1) and (1 >- RPL(2) --> 0.9) and D(1) > 9, then consider VT or VF with very regular waveform; (C) If (RPL(1) --< 1.8 and RPL(2) < 0.9) or (RPL(2) < 1.5 and D(1) < O) or RPL(3) < 0.6, then consider VF. When 0.3 < TR(1) < 0.8, the underlying arrhythmia is VF or VT, and when it is outside this range, it is likely to be a supraventricular rhythm. Conclusions: (A) RPL(j) parameters have a high specificity for discriminating between VT and VF. The method is reliable and simple. (B) The TR( 1) parameter together with RPL(j) allow discrimination between supraventricular tachycardias and ventricular originated tachyarrhythmias. (C) Further analysis must be done using problem-oriented arrhythmias data bases.

r e s u s c i t a t i o n . ~ The m a j o r i t y of s u d d e n d e a t h s are d u e to VF a n d t a k e p l a c e o u t s i d e t h e h o s p i t a l , u s u a l l y in t h e h o m e . 2 Thus, the n e e d for a n a u t o m a t i c defib r i l l a t o r for use in such c i r c u m s t a n c e s has e n c o u r a g e d t h e d e v e l o p m e n t of a s y s t e m to sense t h e prese n c e o f VF. Use o f a u t o m a t i c defibrillation occurs in t w o different clinical contexts: i n t e r n a l l y i m p l a n t e d devices for v e r y h i g h - r i s k patients, a n d e x t e r n a l devices for

The d e t e c t i o n of v e n t r i c u l a r fibrillation (VF) a n d c o r r e c t i o n b y electrical c o u n t e r s h o c k , at t h e earliest t i m e after its o n s e t is essential to e n s u r e successful From the Departament of*Electronics and flnformatica, Universidad Polit~cnica de Valencia, Valencia, Spain.

Supported by the Comision Interministerial de Ciencia y Technologie, Grant PA86-124. Reprint requests: Sergio Guill~n, EE, PhD, Departamento de Electr6nica, Universidad Polit~cnica de Valencia, Camino de Vera, 14, 46020 Valencia, Spain.

253

254

Journal of Electrocardiology Vol. 22 Supplement

emergency use outside the hospital. Mirowsky et al. 3 developed an automatic implantable defibrillator (AID) and performed the first human implant in 1980. The option of using an AID is appropriate for a small number of patients at high risk for VF who have not responded to conservative treatment. However, AID has experienced an impressive evolution. The use of portable automatic external defibrillators (AED), on the other hand, with automatic detection of cardiac arrest rhythms outside the hospital, should enhance the early detection and correction of VF and thus improve the patient's chance of survival in more widespread circumstances. In both cases, reliable automatic VF detection is of crucial importance. VF must be accurately identified and discriminated from ventricular tachycardia (VT) and other ventricular and supraventricular arrhythmias (SVT). Analysis of VF signals and other threatening arrhythmias can be made either in the frequency or in the time domain. In the first case, analysis is accomplished using power spectrum analysis. 4 Fast Fourier Transform (FFT) analysis requires extensive computation and rather complex criteria for detecting VF, since there is usually considerable variability in the spectra of different subjects, producing false positives due to misinterpretation of VT.5 In the time domain, there are many approaches. These techniques include, among others, the analysis of sine content of the VF signal, 6 the peak-and-valley analysis, ~ and the probability density function analysis. s The latter has been used in the AID's algorithms for VF detection; however some problems in

T a b l e 1. A H A - A D B R e c o r d s : M a i n D a t a

Tape

File Name

Record

Reference Bit

Delay (sec)

8201 8201 8201 8201 8201

8201N1 8201N2 8201N3 8201N4 8201N5

ahalnlel ahaln2el ahaln3el ahaln4el ahalnSel

1449 1449 1449 1449 1449

2 10 40 70 120

8202 8202 8202 8202 8202

8202N1 8202N2 8202N3 8202N4 8202N5

aha2nlel aha2n2el aha2n3el aha2n4el aha2n5el

1680 1680 1680 1680 1680

1 35 90 40 67

8203 8203 8203 8203 8203 8203

8203N1 8203N2 8203N3 8203N4 8203N5 8203N6

aha3nlel aha3n2el aha3n3el aha3n4el aha3n5el aha3nn6el

1564 1564 1564 1564 1564 1564

1 54 59 180 420 660

distinguishing VF from VT, ventricular flutter, and other rhythms have been reportedY '~~ Also in the time domain, the autocorrelation function (ACF) has been used to study the VF signal, ~1,z2 but the ACFs of VF signals were not compared with those of other arrhythmias. More recently, Chen et al. ~3 proposed a regression test on the ACF for VF detection. Their analysis comprises two steps: (1) calculation of the ACF; (2) regression test on a plot of peak magnitudes of the ACF against lag values. A diagnosis of VF is given when the ACF/Iag plot of VF does not pass a linear regression test. The objective of our work is to develop a method for detecting VF and other life-threatening arrhythmias, based on direct and simple peak analysis of the ACF. This method may differentiate between fibrillating and nonfibrillating rhythms and in the first case between "coarse" and "fine" VF.

Material and Methods Arrhythmia Records Two sources of ECG records were used in the development of this work. 1. A special-purpose arrhythmia data base (Local Arrhythmia Data Base: LADB), consisting of ECG records obtained from cardiac surgery patients (n = 110) under cardiopulmonary bypass (CPB). The ECG signals were recorded with the collaboration of the cardiovascular surgery service of the hospital "La Fe." ECGs (leads I, II, and III) were recorded on magnetic tape during normal rhythm (NR), VT, and VF. Long tapes were analyzed off-line and 4-sec segments (short records) were digitized at a sampling frequency of 200 Hz, then stored on magnetic media for further processing. 2. The American Heart Association Arrhythmia Database (AHA-ADB). In this case we also selected 4-sec segments from tapes 8201, 8202, and 8203 to make up the VF group. The original sampling frequency is 250 Hz. Segments were chosen from a few seconds before the onset of VF, during onset, and during the evolution of the arrhythmia. Therefore, episodes of VT induced by premature ventricular contractions (PVC), sustained VTs, and ventricular flutter, as well as VF were obtained. Due to the AHAADB characteristics we were able to pick up highamplitude ("coarse") and low-amplitude ("fine") VF episodes. Table 1 summarizes the corresponding data.

Autocorrelation Function Peak Analysis Theoretical Background

Time series ACF is defined as a new time series that, in general terms, can be calculated as~4: Autocovariance (k) Variance

R(k) =

(1)

For practical purposes: ((z(t)

R(k) = ' = ] : N - k

-

IX) * ( z ( t

+

(z(t) -

~,

k)

-

#,))

Ix)

I:N

I=

k=0,

l .....

K

(2)

T@*m*

.a

i

255

Guillen et al.

where k is the lag at the autocorrelation t e r m calculated, N is the total sample number, and p, is the m e a n value of the series. R(k) values fall in the interval ( - 1, 1). W h e n the original function is periodic, it is reflected in the shape of the ACF whose peaks a p p e a r at the same interval of the signal period and w h o s e amplitudes lie on a straight line given by R ( k ) = 1 - k/N. W h e n the period and shape of the signal b e c a m e unconstant and irregular, the ACF peaks decrease more quickly, reflecting those changes. By definition, the ACF of white noise is 0 f o r k - > 1. These general observations can be applied to the particular case of ECG signals. In VT, w h e n the period is rather constant and the shape resembles a sinusoidal wave, the ACF peaks decrease continu-

I)*l

~ 4

*

.

.

.

.

U

.

.

.

.

!

9 eP .

.

A ** .

p)L o

.

U

.

.

.

.

m

.

.

.

.

m~ 7

'i"

9 9 9

.4

I .I ~ l ~

I[~s s mm~md

S .l

m o.

I 90

.4

.o

^~

I.m I

~lm*


i ~ a~lq

d

1.0

c

9

.

.

t.o

l ~ l 1o

11~4

I

.

.

.

.

!

.

.

.

.

8ha u

9 .

.

u

.

.

.

.

i ,-Ihl

A

-

II

aN

~

Imggl I

"rime

. .0

.

.

.

I *4

.

.

.

.

I

9

.

I

.

.

P2

.

.o

I

|.m

...........

lmhlO 8

9

.

.

.

.

i

Soil

.

.

.

Ventricular tachycardia (VT). (Upper panel) ECG segment. (Lower panel) Autocorrection function (ACF). (B) Ventricular fibrillation (VF) with regular waveform. (Upper panel) ECG segment. (Lower panel) Autocorrelation function (ACF). (C) Ventricullar fibrillation (VF) with very irregular waveform. (Upper panel) ECG segment. (Lower panel) Autocorrection function (ACF). F i g . 1. (A)

.


Lo z mleao

256

Journal of Electrocardiology Vol. 22 Supplement

ously, while for VF, the ACF peaks change depending upon the level of "disorganization" of the wave (Fig. 1). Based on these preliminary assumptions, we attempted to quantify such differences through peak analysis of the ACF. With this objective, we have defined the following functions. 1. Positive peak P(j): the highest value determined by the ACF between two zero crossings. Peaks are numbered in their natural order: P(1), P(2) . . . . P(j), and they have three associated parameters: (1) Their corresponding lag kP(j); (2) The time of occurrence TP(j) = kP(j)/sampling frequency; and (3) the time elapsed between the preceding and succeeding zeroes, called the "peak width" PW(j). 2. Control level CL2(k): CL2(k) = 2SE(k), where SE is the standard error of the autocorrelation. Since we can only estimate the autocorrelations, we need to know the error associated with R(k) to determine the extinction of R(k). This value is calculated by the Bartlett approximation ~s : (3)

SE(k) = V'(1/N (1 ___ 2

~

R(v)))

v=l:k

3. Relationship between the actual peak value and the corresponding control level RPL(j): (4)

RPL(j) = P(j)/CL2(kP(j))

4. Monotonic decreasing factor D(j):

(5)

O(j) = RPL(j) - RPL(j + I)

Methods Autocorrelations were calculated over the full length of every segment (800 samples), with k ranging from 1 to 400. A total of 33 short records from the LADB library were analyzed, as were 16 short records from the AHA-ADB. Records from the LADB library were preclassified into three groups: (1) VT and ventricular flutter (group 1); (2) VF with very regular waveform shapes (group 2); and (3) VF with more irregular and chaotic waveforms (group 3). Group 2 comprises those records with an identifiable constant period, while the group 3 records correspond to the ones that do not. Figures 1A, 1B, and 1C illustrate a typical record of each group, with the corresponding ACF, Records belonging to every tape in the AHA-ADB library were analyzed separately. This is because each short segment represents a stage of evolution of the fibrillation, from the onset to the end. In all

cases, the onset was assumed to be as originally marked. Table 1 shows the actual delays of each short segment. In all cases ACFs were analyzed using the corresponding graphics and numerical data, and the above-defined functions were calculated.

Results Analysis of the LADB Records Table 2 summarizes the results. RPL (1) and RPL(2) seem to be sensitive to the differences among the three groups: in group 1, RPL(1) RPL(2) cept in in group 2, RPL(1) RPL(2) cept in in group 3, RPL(1) RPL(2)

(1.78 _+ 0.29) is always > 1 (1.18 __ 0.14) is always > 1 exone case (1.41 __ 0.22) is always > 1 (0.72 _ 0.26) is always --< 1 exone case (0.87 ___ 0.30) is sometimes > 1 (0.49 ___ 0.24) is always < 1

In Figure 2, their respective range of variation are represented as mean value + 2 SD. For VF, group 2 and group 3 were combined. Even though a clear difference in mean values can be observed between VF and VT, wide overlapping bands prevent us from extracting a criterion for discriminating between them. However, if we consider the behavior of RPL(1) and RPL(2) together, a clearer difference is observed a m o n g the groups. Figures 3A and 3B represent the scatterplot of RPL(1) versus RPL(2) and RPL(2) versus RPL(3), respectively. In the first case, the plot area domain has been split into two fields by a straight line, the upper corresponding to D(1) > 0 and the bottom corresponding to D(1) < O. Three zones can be identified with respect to the point's origin: 1. Zone A--all points belong to the VT group and: (6)

RPL(1) > 1.2 AND RPL(2) > 1 AND D(1) > 0

2. Zone B--points correspond to VT or VF from group 2 (regular waveforms) and: (7)

(1.2 -> RPL(1) --> 1) AND (1 ~ RPL(2) 0.9) AND D(1) > 0

Autocorrelation Function Peak Analysis

9 Guillen et al.

257

Table 2. Results of Analysis of LADB Records NUMREG

NOMREG

P1

P2

P3

P4

P5

CL2(I)

CL2(2)

CL2(3)

CL2(4)

CL2(5)

RPL(1)

RPL(2)

RPL(3)

RPL(4) RPL(5)

0,67 0,67 0,80 0,81 0,58 0,52 0.78 0.57 0.52 0.61

0.48 0.48 0.67 0.72 0.44 0,63 0.57 0,44 0.37 0.48

0.41 0.37 0.54 0.67 0.44 0.39 0.44 0.28 0.26 0.39

0.33 0.30 0.48 0.63 0.31 0.50 0.37 0.26 0.26 0.37

0.31 0.30 0.41 0.57 0.28 0.30 0.30 0.15 0.26 0.33

0.31 0.39 0.43 0.43 0.37 0.39 0.33 0.33 0.35 0.35

0.35 0.46 0.52 0.56 0.44 0,52 0,43 0.39 0.39 0,41

0.39 0.50 0.56 0.63 0.48 0.58 0.46 0,43 0.43 0.44

0.41 0.52 0.61 0.67 0.51 0.61 0.48 0.44 0.44 0,48

0.43 0.54 0.63 0.74 0.53 0.65 0.50 0.44 0.45 0.50

2.16 1.72 1.86 1.88 1.57 1,33 2.36 1.73 1.49 1,74

1,37 1.04 1,29 1.29 1.00 1.21 1,33 1.13 0.95 1.17

1.05 0.74 0.96 1.06 0.92 0.67 0.96 0.65 0.60 0.89

0.80 0.58 0.79 0.94 0.61 0.82 0.77 0.59 0.59 0,77

0.72 0.56 0.65 0.77 0.53 0.46 0,60 0,34 0.58 0.66

0.65 0.11

0.53 0.11

0.42 0.11

0.38 0.11

0.32 0.10

0.37 0.04

0.45 0.06

0.49 0.07

0.52 0.08

0.54 0.10

1.78 0.29

1.18 0.14

0,85 0.16

0.73 0.12

0.59 0.12

ecgh3n2 ecgh3n4 ecgh3n6 ecgh4n2 ecgh4n5 ecgh4n7 ecghSnl ecgh5n4 ecgh6n6 ecgh6n7 ecgh7nl

0.49 0.48 0.56 0.64 0.44 0.36 0.67 0.59 0.52 0.70 0.43 0.53 0.10

0.18 0,18 0.40 0.24 0.20 0.46 0,46 0.22 0.33 0.39 0.30 0.31 0.10

0.20 0.20 0.10 0.28 0.12 0,12 0.37 0.11 0.33 0.22 0.24 0.21 0.09

0.28 0.10 0.00 0.32 0.12 0.24 0.30 0.11 0.35 0.22 0.31 0.21 0.11

0.28 0.00 0.00 0.28 0.10 0.24 0.30 0.11 0.26 0.30 0.33 0.20 0.12

0.30 0.40 0.32 0.52 0.32 0.30 0.46 0.44 0.39 0.39 0.37 0.38 0.07

0.34 0.42 0.40 0.56 0.34 0.36 0.52 0.48 0.43 0.44 0.43 0.43 0.07

0.34 0.43 0.40 0.60 0.34 0.36 0.56 0.48 0.46 0.46 0.44 0.44 0.08

0.34 0.44 0.40 0.62 0.36 0.37 0.57 0.48 0.50 0.46 0.46 0.45 0.08

0.34 0.44 0.40 0.40 0.36 0.38 0.59 0.48 0.52 0.48 0.48 0.44 0.07

1.63 1.20 1.75 1.23 1.38 1.20 1.46 1.34 1.33 1.79 1.16 1.41 0.22

0.53 0.43 1.00 0.43 0.59 1.28 0.88 0.46 0.77 0.89 0.70 0.72 0.26

0.59 0.47 0.25 0.47 0.35 0.33 0.66 0.23 0.72 0.48 0.55 0.46 0.15

0.82 0,23 0.00 0,52 0.33 0.65 0.53 0.23 0.70 0.48 0.67 0.47 0.24

0.82 0.00 0.00 0.70 0.28 0.63 0.51 0.23 0.50 0.63 0.69 0.45 0.17

ecgh3n7 ecgh4n4 ecgh4n8 ecgh4n9 5 ecgh5n2 ecgh5n6 ecgh5n7 ecgh6n2 ecgh6nl2 ecgh7n6 ecgh7n7 ecgh7nl3

0.54 0.36 0.28 0.24 0.44 0.27 0.35 0.12 0.30 0.15 0.22 0.19

0.19 0.10 0.05 0.12 0.23 0.26 0.26 0.12 0.09 0.26 O. 11 0.22

0.14 0.20 0.36 0.20 0.22 0.10 0.07 0.16 0.20 0.09 0.11 0.26

0.11 0.05 0.00 0.16 0.26 0.15 0.19 0.00 0.11 0.15 0.11 0.31

0.00 0.00 0.00 0.24 0.30 0.11 0.22 0.00 0.11 0.28 0.00 0.11

0.37 0.50 0.40 0.30 0.35 0.27 0.35 0.38 0.30 0.26 0.22 0.31

0.41 0.50 0.42 0.32 0.37 0.31 0.35 0.42 0.30 0.30 0.24 0.33

0.41 0.52 0.46 0.32 0.38 0.31 0.35 0.42 0.31 0.30 0.24 0.35

0.41 0.52 0.46 0.34 0.39 0.31 0.37 0.42 0.31 0.31 0.26 0.38

0.41 0.52 0.46 0.36 0.41 0.31 0.39 0.42 0.31 0.31 0.26 0.39

1.46 0.72 0.70 0.80 1.26 1.00 1.00 0.32 1.00 0.58 1.00 0.61

0.46 0.20 0.12 0.38 0.62 0.84 0.74 0.29 0.30 0.87 0.46 0.67

0.34 0.38 0.78 0.63 0.58 0.32 0.20 0.38 0.65 0.30 0.46 0.74

0.27 0.10 0.00 0.47 0.67 0.48 0.51 0.00 0.35 0.48 0.42 0.82

0.00 0.00 0.00 0.67 0.73 0.35 0.56 0.00 0.35 0.90 0.00 0.28

0.29 0.12

0.17 0.07

0.18 0.08

0.13 0.09

0.11 0.11

0.33 0.07

0.36 0.07

0.36 0.07

0.37 0.07

0.38 0.07

0.87 0.30

0.49 0.24

0.48 0.18

0.38 0.24

0.32 0.32

(10)

2. Zone B - - ( 0 . 9 --< RPL(2) _< 1)

Group 1 1 2 3 4 5 6 7 8 9 10

ecgh3n3 ecgh3n5 ecgh3n8 ecgh3n9 ecgh4n3 ecghSn3 ecgh6n8 ecgh6nl 1 ecgh7n2 ecgh7n4

Mean

SD Group 2 1 2 3 4 5 6 7 8 9 10 11 Mean SD

Group 3 1 2 3 4 5 6 7 8 9 10 11 12 Mean SD

3. Zone C - - a l l points correspond to VF records and:

AND RPL(3) > 0.6

(RPL(1) --< 1.8 AND RPL(2) < 0.9) (8)

or RPL(2) < 1.5 AND D1 < 0

In the latter case, all records with RPL(I) < 1 belong to group 3, VF, with more irregular and chaotic waveforms. In Figure 3B, the plot domain is also split into two fields, corresponding to D(2) < 0 in the upper one and D(2) > 0 in the bottom one. The above defined zones are characterized by: (9)

1. Zone A - - R P L ( 2 ) > 1 AND RPL(3) > 0.6

3. Zone C--(RPL(2) < 0.9

(11)

AND RPL(3) < 0.9) or

RPL(3) < 0.6

Analysis of the A H A - A D B Records

Table 3 summarizes the results. Record I of Tape 8201 (AHA1N1E1) corresponds to the transition from a regular rhythm (v waves) to a sustained VT, which

258

Journal of Electrocardiology Vol. 22 Supplement

2.4

'''

*

' ' ' '

I ' ' ' '

I

''

'

'-F T-r~-'

1 '

'

'

'

VT

].g VF

].4

t

§ ~

[1.4

-13.1

,,,

,

0

I,

I I i I i 1 i.--.~ 9

I I i_~,

2

4

,,,

1

3

I,,,

,'1

5

6

PEAK NUI6ER (J) Fig. 2. Range of variation of parameter RPL(j) corresponding to VT and VF records. Mean value _+ 2 SD.

Table 3. Results of Analysis of AHA-ADB Records NUM-

REC

NOMREC

P1

P2

P3

P4

P5

D2(1)

D2(2)

D2(3)

D2(4)

D2(5)

RPL(1)

RPL(2)

RPL(3)

RPL(4)RPL(5)

0.74 0.88 0.88 0.54 0.56

0.57 0.76 0.72 0.38 0.52

0.44 0.60 0.56 0.32 0.28

0.31 0.52 0.52 0.34 0.08

0.18 0.40 0.48 0.54 0.20

0.51 0.56 0.52 0.40 0.44

0.66 0.72 0.64 0.44 0.52

0.74 0.80 0.72 0.48 0.56

0.76 0.86 0.76 0.49 0.56

0.78 0.92 0.80 0.56 0.57

1.45 1.57 1.69 1.35 1.27

0.86 1.06 1.13 0.86 1.00

0.59 0.75 0.78 0.67 0.50

0.41 0.60 0.68 0.69 0.14

0.23 0.43 0.60 0.96 0.35

0.72 0.37

0.59 0.30

0.44 0.23

0.35 0.21

0.36 0.20

0.49 0.24

0.60 0.30

0.66 0.33

0.69 0.35

0.73 0.37

1.47 0.41

0.98 0.25

0.66 0.21

0.51 0.21

0.52 0.60

0.34 0.76 0.72 0.28 0.76

0.32 0.56 0.46 0.14 0.50

0.24 0.40 0.36 0.20 0.44

0.26 0.28 0.36 0.20 0.32

0.00 0.24 0.32 0.20 0.24

0.32 0.52 0.48 0.40 0.56

0.36 0.64 0.56 0.40 0.64

0.40 0.68 0.42 0.42 0.72

0.44 0.70 0.64 0.43 0.74

0.44 0.72 0.68 0.46 0.76

1.06 1.46 1.50 0.70 1.36

0.89 0.88 0.82 0.35 0.78

0.60 0.59 0.86 0.48 0.61

0.59 0.40 0.56 0.47 0.43

0.00 0.33 0.47 0.43 0.32

0.57 0.24

0.40 0.18

0.33 0.13

0.28 0.11

0.20 0.15

0.46 0.16

0.52 0.19

0.53 0.20

0.59 0.21

0.61 0.22

1.22 0.36

0.74 0.25

0.63 0.17

0.49 0.17

0.31 0.24

0.93 0.86 0.67 0.13 0.21 0.28

0.84 0.67 0.52 0.37 0.16 0.33

0.73 0.54 0.35 0.20 0.05 0.22

0.63 0.45 0.26 0.18 0.11

0.52 0.32 0.24 0.05 0.16

0.41 0.41 0.39 0.23 0.28 0.33

0.56 0.48 0.44 0.33 0.30 0.37

0.47 0.54 0.48 0.35 0.30 0.41

0.71 0.58 0.50 0.37 0.30 0,41

0.74 0.60 0.20 0.37 0.32 0.41

2.27 2.10 1.72 0.57 0.75 0.85

1.50 1.40 1.18 1.12 0.53 0.89

1.55 1.00 0.73 0.57 0.17 0.54

0.89 0.78 0.52 0.49 0.37 0.00

0.70 0.53 1.20 0.I4 0.50 0.00

0,51 0.31

0.48 0.24

0.35 0.22

0.33 0.18

0.26 0.14

0.34 0.15

0.41 0.18

0.43 0.19

0,48 0.22

0.44 0.24

1.37 0.64

1.10 0.38

0.76 0.40

0.51 0.26

0.51 0.34

Group I 1 2 3 4 5

ahalnlel ahaln2el ahaln3el ahaln4el ahaln5el

Mean SD

Group 2 l 2 3 4 5

aha2nleI aha2n2el aha2n3el aha2n4el aha2n5el

Mean SD

Group 3 1 2 3 4 5 7 Mean SD

aha3nleI aha3n2el aha3n3el aha3n4el aha3n5el aha3n6el

Autocorrelation Function Peak Analysis

9

G u i l l e n et al.

259

3

2.8 ZONEC

2.6

ZOEA

2.4

~ 13

2.2 ZOIE B

2

%

I,

J

1.8

+..+++.:.+"

1.6-

P

1,4-

:

1.2-

:1:

1-

.o

0.8-

,,

0.6

+

,

o ~

r

:

~

+-- - - -- ~ , , ~

]

ZOUE C

,+..f

0,4 0.2 0

I

I

I

I

0.2

I

0,4

A

a

I

I

I

0.6

W

I

0.8

I

I

1

I

~'

4-

I

1.2

I

I

1,4

,>

I

I

1.6

I

118

2

w(~)

2 1.9-

1.81.7 1.0 1.5

ZONE r

1.3 1.2

0-8

~,

"

"

0.7

o-8 0.50.40.30.2-

+ +

]

~ r

o ~df 0 ~ , v <>

~

o

[

+ +~'~. :+" ".,:,. ,~ . +/+++2 ~ ...................

{1

44-

zo~c

4-

,r

4- O

4-

0

4"

0.1 0

I

0

I

o.2

I

I

o`4

I

I

I

0.5

I

u

I

I

1

I

I

1.2

I

I

1`4

I

I

1.,

I

I

1.8

I

2

Fig. 3. (A) Scatterplot of RPL(1)versus RPL(2). (B) Scatterplot of RPL(2)versus RPL(3).

causes a slight distortion in the ACF waveshape, while records 2 and 3 recover the features of VT. Records 4 and 5 correspond to VF. The first record of Tape 8202 presents values of RPL(1) and RPL(2), which fall outside the range, e v e n though it corresponds to a VT episode with narr o w ventricular v waves. Finally, tape 8203 monitors a process that starts with a high-frequency ventricular flutter and evolves into "coarse" VF and finally into "fine" VF. This evolution can be followed by examing the corresponding RPL(j) values from Table 3. The purpose of analyzing these tapes was to submit o u r preliminary conclusions to a first test with rec-

ords from different origins. The observed deviation from the expected results points to a certain limitation of ACF peak analysis in fully characterizing VT-VF waveforms, which is discussed below. However, results from tape 8203 demonstrate the ability of the ACF parameters RPL(j ) to follow the evolution of VF.

Analysis

of the ACF

Waveform

We h a v e observed that P(1) and P(2) amplitudes are very sensitive to variations of the signal period.

260

Journal of Electrocardiology Vol. 22 Supplement

To highlight this feature, Figure 4 shows the scatterplot of RPL(1) versus RPL(2) of arrhythmia episodes distinct from VF or VT, while original zones A and C are overprinted. Two of a total five points fall into zone C, expressing an important shortcoming of the peak analysis method w h e n it is taken alone, although the total number of samples is not statistically significant. This kind of difficulty may appear during supraventricular tachycardias (SVT) and atrial fibrillation (narrow QRS and irregular period). To overcome this, we analyzed the behavior of the defined variable PW(k), assuming it to be sensitive to the actual main ventricular wave width (ie, the QRS complex width). In other words, the relation between PW(1) and TP(1) is comparable with the relation between the signal period and the width of the ventricular wave. Figure 5 shows the above-mentioned relation TR(1) -- PW(1)/TP(1) for both VF and VT records (TR(1) = 0.45 _+ 0.07) and the same records as in Figure 4 (TR(1) = 0.13 _+ 0.03). From this, we defined a range of variations for TR(1) corresponding to ventricular tachyarrhythmias: (12)

Discussion Parameter

One of the starting hypotheses of our work was to assume that identification of arrhythmias could be accomplished through the behavior analysis of some ACF variables: (1) Relative peak amplitudes, (2) relative peak width with respect to period, and (3) time of P( 1 ). P(j) is related to the signal period stability and to the repeatability of a wave pattern in every period, for the ideal case of a sine wave. We see that the progressive increment in period instability, together with changes in the waveform, produce faster attenuation of the peak amplitudes and significant changes in the ACF. We have considered relative amplitudes with respect to their own SE to avoid the rigidity of a fixed level control. Peak analysis is efficient w h e n signals have a low or null zero content (positive and negative deflections with a width-period relation close to 0.5.) This occurs in VF and in most of VTs. W h e n long isoelectric segments appear, their contribution to the autocorrelation are low or negligible, resulting in narrow peaks whose amplitudes became

0.3 < TR < 0.8

which must be considered together with RPL(j) and heart rate, to detect and identify VF.

f- . . . . . .

Definition

I

A

I I

I I I I

a. I,

t:]

c]

J

1.5 C L_

0.5

!

0

I

I

1

I

2 o

I

i

3

Rrt.(2)

sv mythnw

Fig. 4. Scatterplot of RPL(I) versus RPL(2), corresponding to supraventricular arrhythmias (SV).

Autocorrelation Function Peak Analysis

9

Guili~n et al.

261

U

0.7

0.6

m

0

O

0.5 P v

D

U

OD

DO D

DD

0.4

_0_ D

O

0.2

13

[ID

1313

_

_ tl

. . . . . .

o

O

§

0.1

I

O

I

I

I

I

8

4

D

I

12

YF acd Vl"

I

I

I

I

I

18

20

NUhur.~ +

IP4 r t ~ M w

I

24

I

I

28

I

I

32

Fig. 5. TR(1) from VT, VF, and SVT records. m o r e dependent on period stability. Some VTs, like the ones described in tape 8201 and 8202, or supraventricular tachycardias with narrow QRSs produce this effect. The p a r a m e t e r TR(1) seems to be useful in overcoming this drawback, h o w e v e r some SVTs can present a wide QRS w h e n a bundle branch block is present or w h e n aberrant ventricular conduction occurs. In this case it can be interpreted as VT, and further analysis should be performed to diagnose accurately the actual origin. F r o m the AED perspective, the first level of action is to discriminate b e t w e e n rhythms producing hem o d y n a m i c impairment, such as VT or VF, which could be reverted to a life-compatible r h y t h m with a suitable electric countershock, and others that do not respond or in which the use of electrical therapy is contraindicated, as in SVT and cardiac arrest. ]7 The second level is to discriminate between VT and VF to select the discharge energy and determine w h e t h e r synchronized countershock triggering is necessary. Parameter TR( 1 ) together with heart rate could be useful to resolve the first level of discrimination. Unfortunately, the n u m b e r of samples we have studied is not enough to extract a definite conclusion, due to the limitations of the arrhythmia libraries we w o r k e d with. However, our preliminary results indicate that further study is worthwhile. Discrimination b e t w e e n VT and VF from direct p e a k analysis has been based on the period stability

and repeatability of wave patterns, both reflected in RPL(j)s. It is not easy to define the border between them, as is s h o w n in Figure 2. Fortunately, in this case both rhythms respond to countershock and no further consequences should arise. In the present state of our work, some conclusions can be outlined. 1. The RPL(j) analysis has a high specificity in discriminating between VT and VF. The method is reliable and simple. 2. The TR(1) parameter together with RPL(j) discriminate between SVT- and ventricular-originated tachyarrhythmias. Further studies are needed to evaluate better its sensitivity and specificity. 3. Further analysis must be carried out using problem-oriented arrhythmias databases. In our study, neither the LADB nor AHA-ADB had episodes of SVT suitable to our needs. An arrhythmia library consisting of records obtained during out-of-hospital emergency medical technicians" operation would better fit the requirements of the work.

References 1. Cummins RO, Eisenberg MS, Stuhs KR: Automatic external defibrillators: clinical issues for cardiology. Circulation 73:381, 1986 2. Cummins RO, Eisenberg MS, Bergner Let al: Automatic external defibrillation: evaluations of its role in

262

3.

4.

5.

6. 7.

8.

9.

Journal of Electrocardiology Vol. 22 Supplement

the home and in emergency medical services. Ann Emerg Med 13:52, 1984 Mirowski M: Prevention of sudden arrhythmic death with implanted automatic defibrillators. Ann Intern Med 97:606, 1982 Nygards ME, Hulting J: Recognition of ventricular fibrillation utilizing the power spectrum of the ECG. Comput Cardiol 393, 1977 Nolle FM, Bowser RW, Badura FK et al: Evaluation of a frequency-domain algorithm to detect ventricular fibrillation in the surface electrocardiogram. Comput Cardiol, 1988 Kuo S, Dillman R: Computer detection of ventricular fibrillation. Comput Cardiol 347, 1988 Breekelmans FEM, Duisterhout JS, Van Dam RA: Detection of llife threatening arrhythmias by successive peak-trough series analysis. Comput Cardiol 361, 1980 Langer A, Heilman MS, Mower MM, Mirowski M: Considerations in the development of the automatic implantable defibrillator. Med Instr 10:163, 1967 Bardy GH, Ivey TD, Stewart R et al: Failure of the automatic implantable defibrillator to detect ventricular fibrillation. Am J Cardiol 58:1107, 1986

10. Manz M, Gerckens U, Luderitz B: Erroneous discharge from an implanted automatic defibrillator during supra ventricular tachyarrhythmia induced ventricular fibrillation. Am J Cardiol 57:1985, 1986 11. Herschleb JN, Heethaar RM, Van Der Tweel I e t al: Signal analysis of ventricular fibrillation. Comput Cardiol 49, 1979 12. Aubert AE, Denys BG, Ector H, De Geest H: Fibrillation recognition using autocorrelation analysis. Comput Cardiol 477, 1982 13. Chen S, Thakor NV, Mower MM: Ventricular fibrillation detection by a regression test on the autocorrelation function. Med Biol Eng Comp 241:2249, 1987 14. Box GEP, Jenkins GM: Time series analysis, forecasting and control. Holden Day, 1976 15. Bartllett MS: J Roy Stat Soc B8:27, 1946 16. Sandoe E, Sigurd B: Arrhythmia: diagnosis and management. In Fachmed AG. Verlag f~r Fachmedien, 1984 17. Cummins RO, Eisenberg ME, Hallstrom AP et al: What is a "save" ? : outcome measures in clinical evaluations of automatic external defibrillators. Am Heart J 110:1133, 1985