Computer Methods and Programs in Biomedicine (2004) 74, 95—108
`
`Automatic arrhythmia detection based
`on time and time—frequency analysis
`of heart rate variability
`
`Markos G. Tsipouras, Dimitrios I. Fotiadis*
`
`Unit of Medical Technology and Intelligent Information Systems, Department of Computer Science,
`University of Ioannina, GR 45110, Ioannina, Greece
`
`Received 2 August 2002 ; received in revised form 27 January 2003; accepted 11 February 2003
`
`KEYWORDS
`Arrhythmia detection;
`Heart rate variability;
`Time—frequency
`analysis
`
`Summary We have developed an automatic arrhythmia detection system, which is
`based on heart rate features only. Initially, the RR interval duration signal is extracted
`from ECG recordings and segmented into small intervals. The analysis is based on both
`time and time—frequency (t — f) features. Time domain measurements are extracted
`and several combinations between the obtained features are used for the training of
`a set of neural networks. Short time Fourier transform and several time — frequency
`distributions (TFD) are used in the t — f analysis. The features obtained are used for
`the training of a set of neural networks, one for each distribution. The proposed
`approach is tested using the MIT-BIH arrhythmia database and satisfactory results are
`obtained for both sensitivity and specificity (87.5 and 89.5%, respectively, for time
`domain analysis and 90 and 93%, respectively, for t — f domain analysis).
`© 2003 Elsevier Ireland Ltd. All rights reserved.
`
`1. Introduction
`
`Arrhythmia is a collective term for any cardiac
`rhythm that deviates from normal sinus rhythm.
`Arrhythmia may be due to a disturbance in impulse
`formation or conduction, or both, but it is not al-
`ways an irregular heart activity [1]. Respiratory
`sinus arrhythmia is a natural periodic variation in
`RR-intervals, corresponding to respiratory activity.
`Impulse formation may be sinus or ectopic, the
`rhythm regular or irregular and the heart rate fast,
`normal or slow [2,3]. Therefore, the detection of
`abnormal cardiac rhythms and automatic discrim-
`ination from the normal heart activity became an
`
`*Corresponding author. Tel.: +30-6510-98803;
`fax: +30-6510-97099.
`E-mail addresses: markos@cs.uoi.gr (M.G. Tsipouras),
`fotiadis@cs.uoi.gr (D.I. Fotiadis).
`
`important task for clinical reasons. Most of the
`studies address the detection and identification
`of life threatening arrhythmias and specifically
`ventricular and atrial fibrillation and ventricular
`tachycardia. Various detection algorithms have
`been proposed, such as the sequential hypothe-
`sis testing [4], the multiway sequential hypothesis
`testing [5], the threshold-crossing intervals [6], the
`auto-correlation function [6], the VF-filter [6] and
`algorithms based on neural-networks [7—9]. Time—
`frequency (t — f) analysis [10] and wavelet analysis
`[11,12] have also been used. Recent approaches
`utilize complexity measure [13] and multifractal
`analysis combined with a fuzzy Kohonen neural
`network [14].
`to the
`refers
`Heart
`rate variability (HRV)
`beat-to-beat heart rate alterations. HRV believed
`to be a good marker of the individual’s health
`condition and heart diseases [15]. Therefore, HRV
`
`0169-2607/$ — see front matter © 2003 Elsevier Ireland Ltd. All rights reserved.
`doi:10.1016/S0169-2607(03)00079-8
`
`APPLE 1012
`
`1
`
`

`

`96
`
`M.G. Tsipouras, D.I. Fotiadis
`
`analysis became a critical tool in cardiology for the
`diagnosis of heart diseases. Time domain analysis of
`RR-intervals includes calculation of several common
`statistical indices [16,17] and graphical representa-
`tion of the RR-interval duration signal [18,19]. Fre-
`quency analysis provides the power spectrum den-
`sity (PSD) of the RR-interval duration signal using
`Fourier transform and autoregressive techniques
`[20—25]. t — f analysis is based on the use of short
`time Fourier transform (STFT), time—frequency
`distributions (TFDs) and wavelet analysis [10—12] of
`the RR-interval duration signal. Other approaches
`for the HRV analysis include methods from nonlin-
`ear mathematics and chaos theory, such as fractal
`[26,27] and approximate entropy [27] analysis.
`More specifically in the t — f analysis Wigner-Ville
`(WV) distribution [28,29] and improved forms of
`WV, such as pseudo Wigner-Ville (PWV) [30—33] and
`smoothed pseudo Wigner-Ville (SPWV) [34—36],
`discrete Fourier transform and selective discrete
`Fourier transform [37—40], cone shaped kernel
`distribution [10], Choi-Williams distribution [41]
`and other exponential distributions [42] have been
`used.
`In this paper we explore time and t — f analysis
`of the RR-interval duration signal in order to detect
`arrhythmic segments in ECGs [43]. Selected fea-
`tures from the time domain and t — f analysis are
`
`extracted. Several combinations of those features
`are used for training a set of neural networks. The
`decision is finally obtained using decision rules.
`
`2. Materials and methods
`
`Our analysis is carried out in four stages. First
`a preprocessing procedure is used to extract the
`tachograms from the ECGs. The tachograms are
`segmented into small segments. Each segment
`contains 32 RR-intervals. In the second stage time
`domain or t — f methods are applied to extract
`the corresponding features. In the third stage the
`extracted features are used for training a set of
`neural networks. In the forth stage detection of
`arrhythmic segments is carried out using decision
`rules which are fed with the outputs of the neural
`networks.
`
`2.1. Preprocessing
`
`Preprocessing is carried out in two steps. In the
`first step we extract the tachograms from the
`ECG recordings. The dataset used in our study
`is the MIT-BIH arrhythmia database [44,45]. This
`database consists of 48 ECG recordings. The length
`of each recording is 30 min, which results to a
`
`Table 1 Beat annotations of the MIT-BIH arrhythmia database
`
`Annotation symbol
`
`Meaning
`
`Classification in our case
`
`N
`La
`Ra
`A
`a
`J
`S
`V
`F
`[
`!
`]
`e
`j
`n
`E
`Pb
`fb
`p
`Q
`|
`
`Normal beat
`Left bundle branch block beat
`Right bundle branch block beat
`Atrial premature beat
`Aberrated atrial premature beat
`Nodal (junctional) premature beat
`Supraventricular premature beat
`Premature ventricular contraction
`Fusion of ventricular and normal beat
`Start of ventricular flutter/fibrillation
`Ventricular flutter wave
`End of ventricular flutter/fibrillation
`Atrial escape beat
`Nodal (junctional) escape beat
`Supraventricular escape beat
`Ventricular escape beat
`Paced beat
`Fusion of paced and normal beat
`Non-conducted P-wave (blocked APB)
`Unclassifiable beat
`Isolated QRS-like artifact
`
`Normal
`Normal
`Normal
`Arrhythmic
`Arrhythmic
`Arrhythmic
`Arrhythmic
`Arrhythmic
`Arrhythmic
`Arrhythmic
`Arrhythmic
`Arrhythmic
`Arrhythmic
`Arrhythmic
`Arrhythmic
`Arrhythmic
`Normal
`Normal
`Normal
`Normal
`Normal
`
`a L and R annotated beats are classified as ‘‘Normal’’ because they cannot be detected as arrhythmic using only
`heart rate and HRV.
`b Beats with annotations P and f are considered as normal because pace is not in the interest of this study.
`
`2
`
`

`

`Automatic arrhythmia detection based on time and time—frequency analysis of heart rate variability
`
`97
`
`total of 24 h of recordings with 112 568 RR-intervals.
`All RR-intervals are used except for those close
`to the start or end of each recording, which are
`excluded during segmentation. Each beat in the
`database is annotated with a character annotation
`(Table 1). The RDNN software, which accompanies
`the MIT-BIH database, is used for QRS detection.
`Then the RR-interval duration signal (tachogram)
`can be obtained.
`In the second step the tachograms are cut into
`small segments of 32 RR-intervals each, and each
`segment is characterized as normal or arrhythmic
`using the MIT-BIH beat annotation. This results to
`3456 segments. For each RR-interval the character-
`ization of the second beat is used for its characteri-
`zation. The characterization of the MIT-BIH arrhyth-
`mia database beats as ‘‘normal’’ or ‘‘arrhythmic’’
`is shown in Table 1. A 32 RR-interval segment is char-
`acterized as ‘‘normal’’ if it contains more than 30
`‘‘normal’’ RR-intervals otherwise is characterized
`as ‘‘arrhythmic’’.
`
`2.2. Time domain analysis
`
`We apply time domain analysis on the segmented
`dataset. Time domain analysis results in indices
`and markers obtained from the tachogram, such
`as mean and standard deviation of RR-intervals,
`mean and standard deviation of differences be-
`tween adjacent RR-intervals, difference between
`the longer and the shorter RR-interval and others.
`The standard deviation of all normal-to-normal
`RR-intervals (SDRR) is the simplest feature that can
`be extracted from the tachogram. The root mean
`square of successive differences of all normal-to-
`normal RR-intervals (r MSSD) and the standard de-
`viation of successive differences of all normal-to-
`normal RR-intervals (SDSD) are also widely used
`
`Table 2 Combinations of time domain features
`
`intervals present-
`indices. The percentage of
`ing time duration difference between adjacent
`normal-to-normal RR-intervals greater than 50 ms
`(pNN50) is another HRV characteristic. In many
`studies this percentage is used with different time
`threshold, such as 5 ms (pNN5) or 10 ms (pNN10)
`[16,17].
`We use all possible combinations among the
`above mentioned time analysis features (each com-
`bination contains one, two, three, four, five or six
`features) to create the pattern set for the classi-
`fication stage. This leads to a total of 63 feature
`combinations, which are shown in Table 2, with
`3426 patterns each. In the third stage (classification
`stage) we train a feed-forward back-propagation
`neural network, for each feature combination. We
`use 1426 patterns randomly chosen from the total
`of 3426 patterns as training set. Several neural
`network architectures have been tested and we
`have chosen the one that performs better: N inputs
`(number of features used in the specific combina-
`tion), one hidden layer with 20 neurons and one
`output, being a real number between 0 and 1. The
`final ‘‘normal’’ or ‘‘arrhythmic’’ classification is
`made with 0.5 threshold on the networks’ output.
`The training of the neural network ends when the
`square error is less than 0.01 or the training epochs
`are more than 2000. Finally, we result in a set with
`63 neural networks (one for each combination). The
`outputs of the neural networks are fed into a set of
`decision rules as it is described below (forth stage).
`
`2.3. Time—frequency analysis
`
`STFT and various TFDs are used for the t — f analy-
`sis of the segmented dataset. For STFT, the signal
`x(u) is pre-windowed around a time instant t, and
`the Fourier Transform if calculated for each time
`
`Feature combination
`
`Features
`
`1
`2
`3
`4
`5
`6
`7
`8
`. . .
`60
`61
`62
`63
`
`1
`2
`12
`3
`13
`23
`123
`4
`. . .
`3456
`13456
`23456
`123456
`
`SDNN
`r MSSD
`SDNN, r MSSD
`SDSD
`SDNN, SDSD
`SDNN, r MSSD
`SDNN, r MSSD, SDSD
`pNN5
`. . .
`SDSD, pNN5, pNN10, pNN50
`SDNN, SDSD, pNN5, pNN10, pNN50
`r MSSD, SDSD, pNN5, pNN10, pNN50
`SDNN, r MSSD, SDSD, pNN5, pNN10, pNN50
`
`3
`
`

`

`98
`
`M.G. Tsipouras, D.I. Fotiadis
`
`instant t.
`STFT(t, f) =
`
`(cid:1) +∞
`
`x(τ)h(τ − t)e−ifτdτ
`
`(1)
`
`−∞
`where h(t) is a short time window, localized around
`t= 0 and f= 0. STFT suffers of trade-off between
`its window length and its frequency resolution. The
`TFDs used in our study belong to the Cohen’s class
`of distributions [46] and are given by the following
`formula:
`ρ(t, f) =
`
`(cid:1) (cid:1) (cid:1)
`(cid:2)
`
`x
`
`u + 1
`2 τ
`
`ei2πυ(u−t)g (υ, τ) x∗
`(cid:3)
`
`(cid:3)
`
`(cid:2)
`
`u − 1
`2 τ
`
`e−i2πfτdυdudτ
`
`(2)
`
`where x*(t) is the complex conjugate of the signal
`and g(υ,τ) is an arbitrary function called kernel.
`The kernel is different for each TFD. Table 3 shows
`selected TFDs, which are used in our study and the
`corresponding kernels [46—55].
`
`For each 32 RR-interval segment STFT and all
`TFDs are applied (totally 19 methods) and the
`PSD is computed (Fig. 1). For STFT a Hamming
`nine-point length window is used. All TFDs, except
`Margenau-Hill, Page, Rihaczek and Wigner-Ville,
`use frequency and/or time smoothing windows,
`which were set as Hamming nine-point length win-
`dows. For the computation of the STFT and the TFDs
`for each segment we use the previous and the next
`segment to avoid problems in the margins of the
`segment. The PSD is computed and the amplitude
`is normalized in [−1,1] interval. This represents
`the fractional energy of the signal in time t and
`frequency f (Fig. 1a). We obtain horizontal slices
`from the PSD for amplitude= 0.0, 0.2, 0.4, 0.6,
`0.8 and 1.0, which contain the corresponding PSD
`trace (Fig. 1b). The areas between adjacent slide
`traces are calculated (Fig. 1c and d). These areas
`are the t — f features. Six features for each TFD are
`computed and they are used for the training of the
`neural networks (Fig. 2).
`
`Kernel (g(υ,τ))
`sin(πυτ)
`πυτ
`
`τ τ
`
`1
`
`1
`
`(cid:3)2N(cid:2)
`(cid:3)
`
`υ υ
`
`(cid:2)
`
`sin
`
`h(τ)
`
`1 +
`1
`(cid:2)
`e−(πυτ)2/2σ2
`2πσυ
`|τ|α
`πυ
`Cos(πυτ)
`h(τ)cos(πυτ)
`h(τ)cos(πυτ)A∗
`x(υ, τ),
`ej␲␷|␶|
`h(τ)ejπυ|τ|
`1
`h(τ)
`G(υ)h(τ)
`(cid:4) +∞
`ejπυτ
`(cid:4) +∞
`−∞ h(t)e−j2πυτtdt
`(cid:4) +∞
`−∞ h(t)e−j2πυτtdt
`(cid:4) +∞
`−∞ h(t)e−j2πυτtdt
`−∞ h(t)e−j2πυτtdt
`sin(πυτ)
`πυτ
`
`(cid:3)2M
`
`N,M,υ1,τ1>0
`
`σ: scaling factor
`
`σ: scaling factor; a: dissymmetry ratio
`
`h(τ): window function
`Ax(υ,τ): ambiguity function
`
`h(τ): window function
`
`h(τ): window function
`h(τ): window function
`
`h(t): Bessel window
`
`h(t): Hanning window
`
`h(t): binomial window
`
`h(t): triangular window
`
`h(τ): window function
`
`Table 3 TFDs
`
`Distribution
`
`Born-Jordan
`
`Butterworth
`
`Choi-Williams
`
`1
`
`2
`
`3
`
`Generalized rectangular
`4
`Margenau-Hill
`5
`Pseudo Margenau-Hill
`6
`Margenau-Hill-spectrogram
`7
`Page
`8
`Pseudo page
`9
`10 Wigner-Ville
`11
`PWV
`12
`Smoothed PWV
`13
`Rihaczek
`
`14
`
`15
`
`16
`
`17
`
`18
`
`Reduced interference (Bessel window)
`
`Reduced interference (Hanning window)
`
`Reduced interference (Binomial window)
`
`Reduced interference (Triangular window)
`
`Zhao-Atlas-Marks
`
`4
`
`

`

`Automatic arrhythmia detection based on time and time—frequency analysis of heart rate variability
`
`99
`
`Fig. 1 PSD for RR-interval segments containing: (a) normal sinus rhythm (from recording 100 of the MIT-BIH database);
`(b) normal sinus rhythm with two premature ventricular contractions (from recording 105 of the MIT-BIH database);
`(c) normal sinus rhythm mixed with premature ventricular contractions (from recording 200 of the MIT-BIH database);
`(d) rhythm consisting of premature ventricular contractions, left bundle branch block and right bundle branch block
`beats changing to premature ventricular contractions and then to ventricular flutter/fibrillation (from recording 207
`of the MIT-BIH database).
`
`5
`
`

`

`100
`
`M.G. Tsipouras, D.I. Fotiadis
`
`Fig. 1 (Continued ).
`
`For STFT and each TFD we train a feed-forward
`back-propagation neural network, using 1426 pat-
`terns randomly chosen from the total of 3426
`patterns as training set. The architecture is sim-
`ilar to the one used in time domain analysis: six
`inputs, one hidden layer with 20 neurons and
`
`one output being a real number in the interval
`[0,1]. Finally, we result in a set with 19 neural
`networks (one for each method). The outputs of
`the neural networks are fed into a set of deci-
`sion rules (forth stage), which is common for both
`procedures.
`
`6
`
`

`

`Automatic arrhythmia detection based on time and time—frequency analysis of heart rate variability 101
`
`Fig. 1 (Continued ).
`
`2.4. Arrhythmia detection
`
`• Vote: For each segment all neural networks vote
`(cid:10)
`if it is arrhythmic, with threshold 0.5, i.e.:
`0 if yi ≤ 0.5
`Ai =
`1 if yi > 0.5
`where yi is the output of the ith neural network
`and Ai the vote of the ith neural network. If more
`than half votes are accumulated then the decision
`is ‘‘Arrhythmic’’, otherwise is ‘‘Normal’’, i.e.
`Ai ≤ N
`2
`
`(4)
`
`(5)
`
`Normal(0)
`
`if
`
`T =
`
`
`
`i=1
`• Decision vote: Each neural network ‘‘decides’’ if
`(cid:10)
`it will vote using:
`0
`if
` i =
`yi
`if
`where yi is the output of the ith neural network
`and i the vote of the ith neural network. If all
`neural networks vote 0 for a segment then the
`vote is calculated as:
` i = yi
`
`N(cid:9)
`N(cid:9)
`
`i=1
`
`N 2
`
`Ai >
`
`Arrhythmic(1)
`
`if
`
`(cid:11)(cid:11)yi − 0.5
`(cid:11)(cid:11)yi − 0.5
`
`(cid:11)(cid:11) ≤ 0.1
`(cid:11)(cid:11) > 0.1
`
`,
`
`(6)
`
`(7)
`
`For both time domain and t — f analysis we use the
`remaining 2000 segments of the segmented dataset
`as test set. The outputs from all neural networks
`trained for each approach (63 for time domain anal-
`ysis and 19 for t — f analysis) are fed into the fol-
`lowing decision rules to classify each segment as
`normal or arrhythmic:
`• Average: For each segment we calculate the av-
`erage of the outputs of all neural networks and a
`threshold 0.5 is used for the final decision, i.e.
`
`
`
`
`T =
`
`N(cid:9)
`
`yi
`
`≤ 0.5
`
`(3)
`
`i=1
`N
`
`N(cid:9)
`
`Normal(0)
`
`if
`
`Arrhythmic(1)
`
`if
`
`yi
`i=1
`N > 0.5
`
`where T is the final decision; N, the number of
`the neural networks and yi is the output of the
`ith neural network.
`
`7
`
`

`

`102
`
`M.G. Tsipouras, D.I. Fotiadis
`
`Fig.2(a)TFD,(b)horizontalslices,(c)traces,(d)featuresfort—fmethods.
`
`8
`
`

`

`Automatic arrhythmia detection based on time and time—frequency analysis of heart rate variability 103
`
`Table 4 Results for sensitivity and specificity for neural networks
`
`(a) Trained with time features
`Combination
`1
`2
`12
`3
`13
`23
`123
`4
`14
`24
`124
`34
`134
`234
`1234
`5
`15
`25
`125
`35
`135
`235
`1235
`45
`145
`245
`1245
`345
`1345
`2345
`12345
`
`Sensitivity (%) Specificity (%) Combination Sensitivity (%) Specificity (%)
`74
`62
`6
`85
`47
`60
`86
`16
`74
`66
`77
`76
`26
`74
`76
`60
`86
`126
`77
`80
`76
`74
`36
`73
`77
`69
`77
`136
`79
`75
`77
`75
`236
`72
`76
`74
`44
`1236
`80
`77
`72
`65
`46
`81
`49
`69
`76
`146
`74
`65
`74
`75
`246
`75
`75
`62
`84
`1246
`77
`77
`77
`75
`346
`76
`76
`68
`79
`1346
`79
`77
`76
`71
`2346
`73
`75
`83
`40
`12346
`75
`78
`69
`69
`56
`79
`49
`71
`73
`156
`73
`67
`73
`76
`256
`72
`78
`68
`79
`1256
`75
`79
`75
`77
`356
`76
`73
`69
`81
`1356
`77
`77
`79
`76
`2356
`73
`76
`81
`40
`12356
`78
`78
`69
`67
`456
`76
`53
`70
`72
`1456
`73
`70
`75
`74
`2456
`74
`75
`66
`77
`12456
`80
`72
`76
`75
`3456
`77
`72
`71
`78
`13456
`73
`77
`78
`71
`23456
`75
`74
`123456
`78
`72
`
`(b) Trained with t—f features
`Distribution
`
`Sensitivity (%) Specificity (%)
`
`72
`Born-Jordan
`71
`Butterworth
`76
`Choi-Williams
`74
`Generalized Rectangular
`73
`Margenau-Hill
`74
`Pseudo Margenau-Hill
`80
`Margenau-Hill Spectrogram
`74
`Page
`80
`Pseudo Page
`69
`Wigner-Ville
`70
`PWV
`75
`SPWV
`75
`Rihaczek
`76
`Reduced Interference (Bessel Window)
`75
`Reduced Interference (Hanning Window)
`71
`Reduced Interference (Binomial Window)
`Reduced Interference (Triangular Window) 73
`Zhao-Atlas-Marks
`80
`STFT
`70
`
`74
`78
`73
`80
`76
`76
`82
`84
`84
`76
`84
`82
`80
`79
`72
`76
`76
`78
`73
`
`9
`
`

`

`104
`
`M.G. Tsipouras, D.I. Fotiadis
`
`The average of all votes is used with threshold
`0.5 for the final decision, i.e.
`
`N(cid:9)
`
`
`
`
`Normal(0)
`
`if
`
` i
`
`i=1
`N
`
`N(cid:9)
`
`≤ 0.5
`
`(8)
`
`T =
`
`Arrhythmic(1)
`
`if
`
` i
`i=1
`N > 0.5
`The transfer function used among the layers
`is the hyperbolic tangent sigmoid function. The
`training function is the Levenberg—Marquardt back
`propagation method [55]. We use the mean squared
`error performance function, which measures the
`network’s performance using the mean of squared
`errors.
`
`2.5. Standard spectral analysis
`
`Standard spectral analysis was applied to the
`RR-interval signal and used for arrhythmia detec-
`tion and the results are compared with the pro-
`posed techniques. For each 32 RR-interval segment
`we apply Fourier analysis. We used principal com-
`ponent analysis (PCA) to reduce the number of
`Fourier coefficients. The minimum fraction vari-
`ance for components was set to 3 and 4% that
`lead to the reduction of the number of coefficients
`
`Table 5 Results for sensitivity and specificity us-
`ing (a) time domain analysis combined with decision
`rules (b) t — f analysis combined with decision rules
`(c) spectral analysis
`
`Decision rule
`
`Sensitivity
`(%)
`
`Specificity
`(%)
`
`(a) Time domain analysis with decision rules
`Average
`80.68
`78.18
`Vote
`82.60
`78.43
`Decision vote
`87.53
`89.48
`
`(b) t—f analysis with decision rules
`Average
`87.64
`Vote
`86.84
`Decision vote
`89.95
`Method
`Sensitivity
`(%)
`
`(c) Spectral analysis
`DFT with PCA (3%)
`DFT with PCA (4%)
`Prony (6,4)
`Steiglitz—McBride (2,8)
`
`73.14
`74.74
`72.95
`70.67
`
`88.65
`89.25
`92.91
`Specificity
`(%)
`
`71.55
`71.02
`70.41
`73.93
`
`from 32 to 16 and 5, respectively. A feed-forward
`back-propagation neural network was trained, us-
`ing 1426 patterns randomly chosen from the total
`of 3426 patterns as training set. The architecture
`of the neural network is similar to the one used
`before for time domain and t — f analysis: 16 in-
`puts for the case of 3% minimum fraction variance
`for components in PCA and five inputs in the case
`of 4%, one hidden layer with 20 neurons and one
`output being a real number in the interval [0,1].
`We have also applied autoregressive mov-
`ing average (ARMA) analysis using Prony’s and
`Steiglitz—McBride methods. The obtained parame-
`ters were used for the training of a feed-forward
`back-propagation neural network. 1426 patterns
`randomly chosen from the total of 3426 patterns
`were used as training set. The architecture of the
`neural networks is similar to the one used in time
`domain and t — f analysis.
`
`Fig. 3 (a) ROC curve for decision rules for neural net-
`works trained with time features. (b) ROC curve for deci-
`sion rules for neural networks trained with t — f features.
`
`10
`
`

`

`Automatic arrhythmia detection based on time and time—frequency analysis of heart rate variability 105
`
`Table 6 AUC marker results for decision rules
`
`AUC marker
`
`Time
`
`85.91%
`86.32%
`93.46%
`
`t — f
`
`93.53%
`92.27%
`95.83%
`
`Decision rule
`
`Average
`Vote
`Weight vote
`
`3. Results
`
`The corresponding sensitivity and specificity for
`arrhythmic segment detection for each neural
`network are computed. The results for sensitivity
`
`and specificity, for the 63 neural networks trained
`with time feature combinations and the 19 neural
`networks trained with t — f features, are reported
`in Table 4a and b, respectively. The results for a
`single neural network are not satisfactory (average
`sensitivity and specificity: 74 and 72% for neural
`networks trained with time features and 74 and
`78% for neural networks trained with t — f fea-
`tures). Therefore, a single neural network cannot
`be used for arrhythmia detection. However, we
`have observed the following:
`1) Each neural network results in high sensitivity
`and specificity for signal segments for which the
`output is lower than 0.3 or higher than 0.7. This
`
`Table 7 Comparative results with other studies in the literature
`
`Author
`
`Dataset
`
`Arrhythmia types/Methods
`
`Results
`
`Sensitivity (%) Specificity (%)
`
`Thakor-Zhu Pan [4]
`
`170 records (8 s)
`
`Clayton-Muray
`Campbell [6]
`
`70 extracts (4 s)
`
`Thakor, Natarajah
`Tomaselli [5]
`Clayton-Muray
`Campbell [7]
`Yang, Device,
`MacFariane [8]
`Afonso-Tompkins
`[10]
`
`Khadra,
`AlFahoum
`AlNashash [11]
`
`ECGs from 30 patients
`
`70 extracts (4 s)
`
`2363 ECGs
`
`MIT arrhythmia database
`45 ECGs
`8 NR, 12 VF, 13 VT, 12 AF
`
`AlFahoum-Howitt
`[12]
`
`158 ECGs
`
`Ventricular Fibrillation
`Ventricular Tachycardia
`Ventricular Fibrilation
`Threshold Crossing Intervals
`Autocorrelation Function
`VF Filter Leakage
`Signal Spectrum Shape
`Supraventricular Tachycardia
`Ventricular Tachycardia
`Ventricular fibrilation
`Neural networks
`Atrial fibrillation
`
`46
`67
`77
`53
`
`84
`92
`
`Ventricular fibrillation
`Atrial Fibrillation
`Ventricular Tachycardia
`Normal Rhythm
`Ventricular Fibrillation
`
`91.7
`91.7
`84.6
`87.5
`100
`
`72
`38
`55
`93
`
`59
`93
`
`83.3
`91.7
`92.3
`87.5
`100
`
`Staley arrhythmia database
`
`Ventricular fibrillation
`
`(37 NR, 49 VF, 49 VT, 20 AF) Atrial Fibrillation
`Ventricular Tachycardia
`Normal Rhythm
`Ventricular Fibrillation
`Ventricular Tachycardia
`Normal Rhythm
`
`700 QRS
`(150 VT, 250 VF, 300 NR)
`
`85.7
`95.2
`100
`100
`97.5
`92.5
`92
`92
`96
`80
`98
`99
`100% identification after 7 s
`
`Ventricular Fibrillation
`Ventricular Tachycardia
`Atrial Fibrillation
`All types included in
`MIT database
`Time analysis, t-f analysis
`
`98.3
`95
`98.3
`87.53
`
`89.95
`
`96.7
`99.2
`100
`89.48
`
`92.91
`
`Minami
`Nakajima
`Toyashima [9]
`Zhang-Zhu
`Thakor-Wang [13]
`Wang-Zhu
`Thakor-Xu [14]
`
`170 records (48 s)
`85 VT, 85 VF, 34 NR
`180 records (6 s)
`(60 VF, 60 AF, 60 VT)
`
`This work
`
`MIT arrhythmia database
`
`48 ECGs 30 min length
`
`11
`
`

`

`106
`
`M.G. Tsipouras, D.I. Fotiadis
`
`is because the output of the neural network can
`be viewed as a possibility function that decides
`if a segment is normal (result 0) or arrhythmic
`(result 1).
`2) Each neural network results in low sensitivity
`and specificity for signal segments for which
`the output was close to 0.5 (i.e. in the interval
`[0.5− k, 0.5+ k] with k≤0.2). This is an uncer-
`tainty interval with high error rate.
`3) For all signal segments there is a number of neu-
`ral networks that can detect them correctly with
`output outside the uncertainty interval [0.45,
`0.55] (i.e. k= 0.05). When the uncertainty inter-
`val is larger (i.e. k= 0.1, 0.15 or 0.2) the num-
`ber of neural networks, which detect the same
`number of segments correctly, is reduced and
`for some segments there are no neural networks
`with output outside the uncertainty interval. We
`have used various uncertainty intervals and the
`best choice is [0.4, 0.6] (i.e. k= 0.1).
`
`The above observations lead us to the use of mul-
`tiple identifiers combined with decision rules. The
`results for sensitivity and specificity when we use
`the decision rules on the neural networks’ outputs
`are presented in Table 5. InTable 5 we present
`also the results for standard spectral analysis us-
`ing Fourier analysis and parametric modeling with
`Prony’s method, with numerator order 2 and de-
`nominator order 8, and Steiglitz—McBride method,
`with numerator order 6 and denominator order 4.
`For each decision rule the receiver operating char-
`acteristic (ROC) curve is computed. The ROC curves
`are shown in Fig. 3. Using the ROC curves the area
`under curve (AUC) marker is calculated and the re-
`sults are presented in Table 6.
`
`4. Discussion—conclusions
`
`We have developed an automatic procedure for
`the detection of arrhythmias using heart rate fea-
`tures. The outcome of the method is the classi-
`fication of the ECG signal segments as ‘‘normal’’
`or ‘‘arrhythmic’’. The method is based on time
`analysis and t — f analysis features. If time features
`are chosen their combination lead us to 63 neural
`networks. If t — f analysis is followed then we re-
`sult into 19 neural networks. We have proven that
`a single neural network does not offer satisfac-
`tory results in terms of sensitivity and specificity
`for arrhythmia detection and this imposed the use
`of decision rules, which combine the outputs of
`all neural networks. One of the advantages of the
`method is that the single use of heart rate features
`can lead to the identification of arrhythmic inter-
`
`vals in ECG recordings. This does not depend on
`the type of arrhythmia.
`For the time domain approach the average and
`vote decision rules result in low performance for
`both sensitivity and specificity (81 and 78%, re-
`spectively). This is expected since in both cases
`the neural networks’ outputs inside the uncer-
`tainty interval have high error rate. The sensi-
`tivity and specificity for decision vote are 87.5
`and 89.5%, respectively. Average and vote deci-
`sion rules for the t — f features resulted in 87%
`sensitivity and 89% specificity. The decision vote
`results superior (90% sensitivity and 93% sensitiv-
`ity). The obtained sensitivity and specificity for
`using Fourier analysis followed by PCA are 73.14
`and 71.55% for minimum fraction variance 3,
`74.74 and 71.02% for minimum fraction variance
`4%, respectively. Parametric modeling resulted in
`72.95% for sensitivity and 70.41% for specificity
`using Prony’s method and 70.67% for sensitivity
`and 73.93% for specificity using Steiglitz—McBride
`method.
`Several researchers have addressed the problem
`of arrhythmia detection using ECG features and
`heart rate analysis. A summary of those studies
`with the results obtained in terms of sensitivity
`and specificity is provided in Table 7. Most of the
`authors [4—14] focus their work on the detection
`of some particular types of arrhythmia (ventricular
`fibrillation, atrial fibrillation, ventricular tachy-
`cardia, supraventricular tachycardia). Our method
`compares well since it addresses the arrhythmia
`detection for all types of arrhythmia, using only
`heart rate features. A drawback is that some ar-
`rhythmia types (such as left bundle branch block
`and right bundle branch block beats) cannot be
`detected using only heart rate features. Another
`important feature is that some of the methods use
`a specific dataset so as to be evaluated, there-
`fore, a direct comparison with our method is not a
`simple task. The dataset used for each method is
`given in Table 7. In our case we have used all the
`recordings of the ‘‘standard’’ MIT-BIH arrhythmia
`database to evaluate the proposed method.
`Our study is based on the analysis of the
`RR-interval duration so the proposed method is
`capable of detecting arrhythmia types that pro-
`duce irregularities on the RR intervals, the HRV
`or the heart rhythm. Left bundle branch block
`and right bundle branch block beats do not pro-
`duce such artifacts, therefore, they cannot be
`detected using only heart rate features. Instead of
`excluding these beats from the analysis we clas-
`sify them as ‘‘normal’’ because we wanted the
`method to be as general as possible. Probably this
`‘‘misclassification’’ affects the results but no tests
`
`12
`
`

`

`Automatic arrhythmia detection based on time and time—frequency analysis of heart rate variability 107
`
`were made on this matter so we cannot make any
`comments on this subject.
`Besides the QRS detection and the extraction of
`the RR-interval duration signal, there is no other
`processing of the ECG recording in our study, such
`as P wave or T wave detection which will make
`the process more complicated and time consuming.
`Therefore, noisy data can be analyzed because QRS
`detection algorithms perform well. The exclusive
`use of the RR-interval duration signal leads to a high
`reduction of input and processing data, compared
`with other ECG analysis methods. Moreover, the fi-
`nal decision is not based on a single identifier but on
`the combined results of a set of identifiers. There-
`fore, the system is expected to have high general-
`ization capability. Due to the short processing time
`and the generalization capability of the method we
`believe that the proposed approach can be used
`in real time arrhythmia detection systems. In addi-
`tion, RR-interval duration features can be used for
`the classification of detected arrhythmic segments
`into several arrhythmia types.
`
`Acknowledgements
`
`The authors are grateful to Professor D. Sideris,
`Professor A. Likas and Professor N. Galatsanos for
`useful comments and suggestions.
`
`References
`
`[1] E. Sandoe, B. Sigurd, Arrhythmia–—a Guide to Clinical Elec-
`trocardiology, Verlags GmbH, Bingen, 1991 (Chapter 3).
`[2] L. Goldberger, E. Goldberger, Clinical Electrocardiography,
`The Mosby Company, Saint Louis, 1977 (Chapter 11).
`[3] D.A. Sideris, Primary Cardiology, Grigorios K. Parisianos,
`Athens, 1991 Scientific Editions, in Greek.
`[4] N.V. Thakor, Y.S. Zhu, K.Y. Pan, Ventricular tachycardia and
`fibrillation detection be a sequential hypothesis testing
`algorithm, IEEE Trans. Biomed. Eng. 37 (9) (1990) 837—843.
`[5] N.V. Thakor, A. Natarajan, G. Tomaselli, Multiway sequen-
`tial hypothesis testing for tachyarrhythmia discrimination,
`IEEE Trans. Biomed. Eng. 41 (5) (1994) 480—487.
`[6] R.H. Clayton, A. Murray, R.W.F. Campbell, Comparison of
`four techniques for recognition of ventricular fibrillation of
`the surface ECG, Med. Biol. Eng. Comp. 31 (1993) 111—117.
`[7] R.H. Clayton, A. Murray, R.W.F. Campbell, Recognition of
`ventricular fibrillation using neural networks, Med. Biol.
`Eng. Comp. 32 (1994) 217—220.
`[8] T.F. Yang, B. Device, P.W. Macfarlane, Artificial neural net-
`works for the diagnosis of atrial fibrillation, Med. Biol. Eng.
`Comput. 32 (1994) 615—619.
`[9] K. Minami, H. Nakajima, T. Toyoshima, Real-time dis-
`crimination of ventricular tachyarrhythmia with Fourier-
`transform neural network, IEEE Trans. Biomed. Eng. 46 (2)
`(1999) 179—185.
`[10] V.X. Afonso, W.J. Tompkins, Detecting ventricular fibrilla-
`tion, IEEE Eng. Med. Biol. 14 (1995) 152—159.
`
`[11] L. Khadra, A.S. Al-Fahoum, H. Al-Nashash, Detection of
`life-threatening cardiac arrhythmias using wavelet trans-
`formation, Med. Biol. Eng. Comput. 35 (1997) 626—632.
`[12] A.S. Al-Fahoum, I. Howitt, Combined wavelet transforma-
`tion and radial basis neural networks for classifying life-
`threatening cardiac arrhythmias, Med. Biol. Eng. Comput.
`37 (1999) 566—573.
`[13] X.S. Zhang, Y.S. Zhu, N.V. Thakor, Z.Z. Wang, Detecting
`ventricular tachycardia and fibrillation by complexity mea-
`sure, IEEE Trans. Biomed. Eng. 46 (5) (1999) 548—555.
`[14] Y. Wang, Y.S. Zhu, N.V. Thakor, Y.H. Xu, A short-time mul-
`tifractal approach for arrhythmia detection based on fuzzy
`neural network, IEEE Trans. Biomed. Eng. 48 (9) (2001)
`989—995.
`[15] Task force of The European Society of Cardiology and The
`North American Society of Pacing and Electrophysiology,
`Heart rate variability: sta