TIME-FREQUENCY ANALYSIS OF HEART RATE VARIABILITY FOR NEONATAL SEIZURE DETECTION

Transcription

1 Australas. Phys. Eng. Sci. Med. Vol. 9, No 1, 06 TIME-FREQUENCY ANALYSIS OF HEART RATE VARIABILITY FOR NEONATAL SEIZURE DETECTION MB Malarvili, Mostefa Mesbah and Boualem Boashash Signal Processing Research Laboratory, Queensland University of Technology, Brisbane, Australia Abstract The ECG has been much neglected in automatic seizure detection in the newborn. Changes in heart rate and ECG rhythm are often found in animal and adult patients with seizure. However, little is known about heart rate variability (HRV) changes in human neonate during seizure. Results of ongoing time-frequency research are presented here with the aim to compare the performance of various time-frequency distributions (TFDs) when applied to HRV time series for non-seizure and seizure newborns. The TFDs studied are the Wigner-Ville (WVD), the Spectrogram (SP), the Choi-Williams (CWD) and the Modified B (MBD) distributions. Based on our preliminary results, our current conclusion is MBD outperforms other TFDs in terms of time-frequency resolution, cross-terms suppression and to represent the newborn HRV signals of non-seizure and seizure which are closely-spaced components in the time-frequency domain. Key words: Newborn seizure detection, heart rate variability (HRV), time-frequency distributions (TFDs). Introduction Neonatal seizures leads to the risk of subsequent death, cerebral palsy, mental retardation, epilepsy, attention/hyperactivity disorders, behavioural disturbances and other Central Nervous system (CNS) based disorders that must be carefully assessed [1]. Growing attention is focused on the development of computerized methods to automatically detect newborn seizure based on the Electroencephalogram (EEG). There are a number of techniques available for detecting neonatal seizures utilizing EEG signals in the time [], frequency [3] and time-frequency domains [4]. However, neonatal seizure recognition remains a very challenging task []. Therefore, an additional guide is essential in developing EEG based seizure detection algorithm. In [6], it was found, the heart rate changes being very early, is the initial manifestations of seizure activity, and should be ideal parameters to assist in the determination of seizure onset. This has provided a motivation to investigate ECG signal to make the seizure detection profound. In addition, [6] reported that heart rate changes are an extremely common feature of complex partial seizures. Seizures can cause extreme alteration to autonomic activity. ECG and variation in ECG characteristics are primarily under control of the autonomic nervous system (ANS), providing a sensitive and non-invasive means of detecting alterations in autonomic activity. Early investigations by neurologists using animal models of epilepsy [7], adult [7,8,9] and children [] suggest that paroxysmal changes in ECG including heart rate, alteration in the RR interval and QT interval attributed to clinical seizure activity. Continuous recording of beat-tobeat heart rate in 9 paralyzed infants indicated that heart rate increases during the seizure activity besides rhythmic fluctuation in ECG [11]. The conclusions proposed by neurologists are case studies based on the observation, continuous monitoring of the behavior of ECG and EEG channels simultaneously. The precise relationship between these changes to seizures has not been specifically determined. The HRV is emerging as a major non-invasive tool in diagnostic and monitoring of the autonomous control as it is able to assess the overall cardiac health and the state of the ANS [1]. The ANS has sympathetic and parasympathetic components. The separate rhythmic contributions from sympathetic and parasympathetic autonomic activity modulate the heart rate, and thus the RR intervals of the QRS complex in the ECG, at distinct frequencies. Sympathetic activity in newborn is associated with the low frequency (LF) range ( Hz) while parasympathetic activity is associated with the higher frequency (HF) range ( Hz) of modulation frequencies of the heart rate. The mid frequency (MF), centres near 0.1 Hz, is both parasympathetically and sympathetically mediated. The HF corresponds to the respiratory and the LF mediated by a variety of different influences [13]. The HRV characteristics have been investigated with different algorithms based on time or frequency domain. Difficulty encountered in frequency domain processing is the large non-stationary behaviour of heart beats. Even for a normal healthy person, the heart beat tend to be time variant. This is because the inter beat interval of the heart rhythm varies markedly due to irregularities in the initiation of the cardiac impulse in the atrium. These nonstationarities become more severe in abnormal cardiac rhythms. Time-frequency (TF) analysis is able to provide the time-local frequency description of HRV necessary to characterize such changing autonomic regulation [14]. The TFD of a signal describes how its spectrum changes with time. There are different techniques to estimate the TFD and to differentiate the HRV for seizure 67

2 Australas. Phys. Eng. Sci. Med. Vol. 9, No 1, 06 and non-seizure newborns. Thus, the aim of this paper is to select the best TFD that characterize the changes in HRV during seizures as an additional guide that may predict or explain the risk of seizures in the newborn. Time-frequency distribution Time-frequency analysis provides time-localized spectral information for a non-stationary signal as a distribution function in terms of time and frequency. In order to achieve fine simultaneous time-frequency resolution in a non-stationary time series, we must deal with the uncertainty principle [1]. The uncertainty principle restricts us from achieving arbitrarily fine resolution simultaneously in both the time and the frequency domain. A trade-off between them should be considered in order to reach the desired "good" resolution. In the majority of the cases, it is dependent on the signal characteristics. Many different types of TFD functions have been developed for the purpose of analysis. The general expression for the quadratic TFD is [16]: ρ ( t, f ) = z e jπ ( νt νu fτ ) τ τ g( ν, τ ) z( u + ) z * ( u ) dνdudτ where z(t) is the analytic associate of the real signal s(t). The function g ( ν,τ ) is known as the TFD kernel filter which determines the characteristics of TFDs and determines how the signal energy is distributed in the time-frequency plane. The general requirements of an appropriate TFD are non-negativity, realness, time and frequency marginals, instantaneous frequency, group delay, time and frequency supports, and time and frequency shift properties [1]. A brief description of WVD, SP, CWD and MBD are given below. The WVD is the fundamental quadratic TFD whose kernel filter is g ( ν, τ ) = 1 [16]. The CWD has a D Gaussian kernel filter, τ σ ( ν τ) ν / g, =e [16]. The CWD has one parameter, σ, which allows the user to select the amount of filtering in the Doppler-lag ( ν, τ ) domain. The SP, whose kernel filter is τ jυu τ g ( ν, τ ) = w * ( u ) e w( u + ) du, performs a local or short-time Fourier analysis by using a sliding analysis window w(u) to segment the signal into short sections centered near the observed time t before computing the Fourier transformation. High resolution in TF plane is possible with other TFDs, such as the MBD. The kernel for MBD g ν, τ = Γ β + jπν / Γ β where Γ (.) stands for the is ( ) ( ) ( ) gamma function and β is a real, positive number that controls the trade off between components resolution and cross-terms suppression [17]. Quantification of heart rate variability The one channel newborn ECG was recorded simultaneously with channel EEG. The EEG was labelled as either seizure or non-seizure by a neurologist from the Royal Women s Hospital, Brisbane, Australia. Figure 1 shows the steps involved in generating the HRV from the ECG signal. A QRS detection algorithm is implemented to extract the R point of the ECG. Conventional time-domain methods, like used in [6, 18] are based on fiducial point estimations of the ECG signal, differentiation to enhance the peaks in the ECG signal and rule-based thresholding to identify the ECG parameters. However, the authors in [19] reported that these methods lead to inaccuracies in the ECG parameters identification and detection and at certain cases misses the event of Q wave which is very crucial in the heart rate determination. Figure 1. Steps for generating the HRV. Thus, we implemented the smoothed nonlinear energy operator (SNEO) to extract the R point which can be thought of as spikes in ECG signal. More information on SNEO can be found in []. The time series of RR interval is called tachogram. The instantaneous heart rate (IHR) is the inverse of the RR interval and shows the variability of heart rate. The unevenly sampled RR interval is interpolated using cubic splines and downsampled at Hz. Finally, the trends of the time series were computed and removed in order to remove artefacts which would mask the frequencies of interest. This series constituted the HRV signal used in the analysis. Figure shows examples of IHR of the nonseizure and seizure babies. Performance analysis ECG Recording QRS Detection RR interval Resampling to Hz HRV As stated in section 1, the time-frequency analysis was focused on four generally used TFDs. Those are the 68

3 Australas. Phys. Eng. Sci. Med. Vol. 9, No 1, 06 Figure. The IHR of (a) non-seizure (b) seizure neonate. Amplitude Amplitude LF: Hz MF: Hz Power Spectrum of HRV for Nonseizure HF: Hz frequency (Hz) LF: Hz MF: Hz Power Spectrum of HRV for Seizure HF: Hz frequency(hz) Figure 3. The frequency domain representations ( FFT ) of the HRV of (a) Figure (a), (b)figure (b). SP, the WVD, the CWD, and the MBD. Because of the space limitation, we present and discuss our TFD performance analysis using only two (1 non-seizure and 1 seizure) out of the 8 neonates studied. These can be considered as representatives of the general characteristics observed. In Figure 3, the frequency domain representations ( FFT ) of the HRV of Figure are shown. Three frequency band of interest which are the HF, MF and LF have been labeled in the power spectrum. The TFDs of HRV for both the non-seizure and seizure signals in Figure are shown in Figures 4 and, respectively. All the plots shown in figures were made using the same plot routine: the left plot represents time series of HRV, the bottom plot represents power spectrum and the center figure shows the joint TFD. The sequence of plots labeled with (a), (b), (c), and (d) corresponds to the TFDs of the WVD, SP, CWD, and MBD, respectively. The relevant frequency band are labeled with LF, MF, HF, and indicated with arrows. For simplicity, the labels are shown only on Figures 4(a), and (a). Because the relative position of those frequencies prevails in all the sequence of figures, only the arrows are indicated in the figures labeled (b), (c), and (d). In Figure 4 (a), the dominant frequency content can be observed in the LF and MF. However, the distribution also shows the hard-to-interpret components around the HF components, and even the dominant components cannot be clearly identified due to the presence of interference terms. In other words, the WVD has been degraded by cross terms because the HRV is a multi timefrequency component signal. The WVD is the optimal TFD for representing monocomponent linear FM signals. However, for nonlinear FM and multicomponent signals, it suffers from artifacts which corrupt the information in the time frequency plane. In Figure 4(b), at least well-defined frequency components can be observed in the LF and HF. Even though the frequency resolution is fairly satisfactory, the SP is lack in time resolution which makes the TF components smeared. The quadratic SP smoothes away all interference terms except those which occur when two signal components overlap. This smoothing has the side effect of reducing signal components resolution. The SP poorly represents rapidly changing spectral characteristics, and it cannot optimally resolve closely spaced components because of the inherent tradeoff between time resolution, which requires a short analysis window, and frequency resolution, which requires a long analysis window. For CWD, the parameter σ of its kernel was set equal to 0.4. It can be seen that many of the cross-terms have been eliminated compared to WVD. However synchronisation effects (i.e., the horizontal lines) prevail. This is due to the trade-off between suppression of the crossterms and the auto-component terms. This also makes the component in MF smeared. Because the CWD kernel has hyperbolic isocontours in the ambiguity plane, it always passes the cross-terms that originate from 69

4 Australas. Phys. Eng. Sci. Med. Vol. 9, No 1, 06 0 LF MF a.) WVD HF 0 c.) Choi William x x 4 0. b.) Spectogram d.) Modified B-distribution x x 4 0. Figure 4. TFD for non-seizure HRV: (a) WVD (b) SP (c) CWD (d) MBD. interaction between signal components that occur at either the same time or same frequency. As for MBD, the parameter, β of its kernel was set to It is noticed that the cross-terms are smaller than observed in the CWD and WVD. It also has better time resolution compared to SP. This improvement facilitates the identification/interpretation of the frequency components of the HRV in non-seizure neonatal. The dominant frequency content can be observed in the LF, MF and HF band. It gives a good estimation of the IF law of each component even at points when the IF law changes rapidly. MBD has high resolution, effective cross-terms reduction and a kernel that does not perform filtering in the lag direction. Numerical analysis showed that this TFD is of superior performance in reducing cross-terms while keeping high-resolution. Results of the TFD analysis of the HRV for seizure baby are presented in Figure. A similar pattern is observed regarding the increase/reduction of the timefrequency resolution, the suppression of cross-term interference and synchronisation effects, as was discussed for the non-seizure HRV. It is worth noting that the energy concentration in the frequency band especially in the HF has reduced for the seizure case. This could be due to the manifestation of the seizure in the autonomic system and as a result, it is noticed in the TFD of HRV which was not evident in the time series of HRV and the respective power spectrum. 70

5 Australas. Phys. Eng. Sci. Med. Vol. 9, No 1, 06 a.) WVD b.) Spectogram x x c.) Choi-William x d.) Modified B-distribution x 4 1 Figure. TFD for seizure HRV: (a) WVD (b) SP (c) CWD (d) MBD. To be a suitable tool for practical high resolution TF analysis, a TFD must be real, preserve the component energy, and reduce the cross-terms, while preserving the resolution and reveal the IF law of a monocomponent signal by its peak. The definition would be valid for every slice of TFDs, taken at time t = to [17]. Thus, for each TFD for the non-seizure case, a normalized slice at the middle of the time interval, t = is taken and displayed as a comparison to the MBD in Figure 6 to look into the TFDs s performance. Figure 6 shows the normalized slices of TFDs plotted in Fig. 4. (a), (b) and (c). From Figure 6 (a), it is clear that the WVD has negative values and the cross-terms oscillate in time between signal components. The SP shows a better performance compared to WVD but still the energy concentration for each component is not preserved. Meanwhile, the CWD failed to exhibit a good suppression of any undesirable artifacts for each of the components. The MBD is found to be the closest to the ideal compromise for the class of signals considered; it is almost cross-terms free and has high components resolution in the TF plane. 71

6 Australas. Phys. Eng. Sci. Med. Vol. 9, No 1, 06 Figure 6. Normalized slices (dashed) of (a) WVD. (b) SP. (c) CWD, all compared against the MBD (solid). Conclusions Among the four different types of Cohen s class discussed above, it appears that MBD is the most suitable for characterizing the property of HRV for neonatal seizure detection. It is found that MBD outperforms other existing distributions in terms of time-frequency resolution, as well as cross-terms suppression and to represent the HRV signals which are closely-spaced components in the time-frequency domain. These are the results of our ongoing study and future research will be focused on the study of the features of the HRV utilizing MBD and characteristics of the seizure and non-seizure in the TF plane in order to provide new insight into newborn seizure detection. References [1]. American Epilepsy Society Medical Education Program, Resident s Version, Appendix A of Clinical Section, page c-6-c- 7. A publication of the American Epilepsy Society, address []. A. Liu, J.S. Hahn, G.P. Heldt, and R.W. Coen, Detection of neonatal seizure through computerized EEG analysis, Electroenc. And Clinical Neurpphysi., vol.8, pp.-37, 199. [3]. J. Gotman, D. Flanagan, B. Rosenblatt, A. Bye and E.M. Mizrahi, Evaluation of an automatic seizure detection method for the newborn EEG, Electroenc. And Clinical Neurpphysi., vol.3, pp: , [4]. H. Hassanpour, M. Mesbah and B. Boasash Comparative Performance of time-frequency based newborn EEG seizure detection using spike detection, International Conference on Acoustics, Speech, and Signal Processing, (ICASSP '03), vol., pp ,April03. []. H. W. Taeusch, R. A. Ballard, C. A. Gleason, Avery s diseases of the newborn. 8 th Edition. Elsevier, USA 0. [6]. S.R. Quint, J.A. Messenheimer, M.B. Tennison, H.T. Nagle, Assessing autonomic activity from the EKG related to seizure onset detection and localization, Second Annual IEEE Symposium on Computer-Based Medical Systems Proceedings, pp: 9, June1989. [7]. S.J. Tavernor et al, ECG QT lengthening associated with epileptiform EEG discharges- a role in sudden unexplained death in epilepsy? Seizure, pp: 79-83, [8]. F. Leutmezer, C. Schernthaner, S. Lurger, K. Pötzelberger, and C. Baumgartner, Electrocardiographic Changes at the Onset of Epileptic Seizures, Vol.44, Issue3, pp:348,epilepsia,march03. [9]. M. Zijlmans, D. Flanagan, J. Gotman, Heart Rate Changes and ECG Abnormalities during Epileptic Seizures: Prevalence and Definition of an Objective Clinical Sign, Vol. 43, Iss.8, pp: 847, Epilepsia, August 0. []. M.E. O'Regan, J.K. Brown, Abnormalities in cardiac and respiratory function observed during seizures in childhood, Developmental Medicine and Child Neurology [NLM - MEDLINE], Vol.47, Iss. 1; pg. 4, Jan 0. [11]. R.N. Golberg et al, Detection of seizure activity in the paralysed neonate using continuos monitoring, Pediatrics, vol 69, No [1]. M.V. Kamath, Time-frequency Analysis of Heart rate Variability signals in Patients with Autonomic Dysfunction, pp: , TFTS 96. [13]. J.P. Finley, S.T. Nugent, Heart rate variability in infants, children and young adults Journal of Autonomic Nervous System, Vol.1; pp:3-8, 199. [14]. R.M.S.S. Abeysekera, Time-Frequency Domain features of electrocardiographic signals: An Interpretation and their application in computer aided Diagnosis University of Queensland, PhD thesis, [1]. L.Cohen, Time-Frequency Distribution-A Review, Proc. Of IEEE, Vol.77, pp. 941:980, July [16]. B. Boashash. Time frequency Signal Analysis and Processing: A Comprehensive Reference, Oxford, UK: Elsevier, 03. [17]. B.Boashash, V. sucic, Resolution Measure criteria for the Objective assessment of the Performance of Quadratic Time- Frequency Distributions, IEEE Trans. On Signal Processing, Vol.1, pp:13-163, May 03. [18]. 13. R. D. Ranganath et. al ECT and Heart Rate Changes: An Alternative to EEG Monitoring for Seizure Confirmation During Modified ECT, German Journal of Psychiatry, pp:31-34, 03. [19]. T. Srikanth, S. A. Napper, H. Gu. Bottom-Up Approach to Uniform Feature Extraction in Time and Frequency Domains for Single Lead ECG Signal, International Journal of BioElectromagnetism, Vol.4, No.1, 0. []. MB Malarvili, H. Hassanpour, Mostefa Mesbah and B. Boashash, A Histogram-Based Electroencephalogram Spike Detection, ISSPA 0. 7