Bimodality and Scaling in Recurrence Networks from ECG data

G. Ambika; Sneha Kachhara

REVIEW 4 major objections 6 minor 20 references

Bimodality and Scaling in Recurrence Networks from ECG data

T0 review · 4 major / 6 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read This paper claims that one-minute ECG recordings, turned into recurrence networks, show a two-peaked degree distribution and two distinct scaling regimes in link density, with disease-specific values.

desk verdict A promising but overclaimed proof-of-concept: the bimodality and scaling observations are new, yet the paper's own text shows the headline features are threshold-dependent and not class-separating as stated. read the letter →

arxiv 1908.01286 v1 pith:7BNN6L53 submitted 2019-08-04 physics.med-ph nlin.CDphysics.soc-ph

Sneha Kachhara , G. Ambika This is my paper · ORCID

classification physics.med-phnlin.CDphysics.soc-ph PACS 05.45.Tp64.60.aq05.45.-a

keywords recurrencenetworksECGdegreedistributionbimodalitylinkdensityscalingexponentsphasespaceembeddingcardiacdynamics

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper sets out to show that a one-minute clinical ECG recording, when converted into a recurrence network, carries two stable structural signatures of cardiac dynamics: a degree distribution with two separated peaks, and a link density that scales with the recurrence threshold in two distinct regimes. The authors report that both features appear across all ECG classes they study, while networks from standard chaotic, hyperchaotic, and noise signals are unimodal and scale differently. They interpret the two peaks as evidence of two spatial scales in the heart's reconstructed phase-space attractor, and the two scaling exponents as finer quantifiers of that structure. If the claim holds, recurrence networks give a way to quantify cardiac complexity from short, nonstationary clinical data using the full ECG waveform, not just heart-rate variability.

What carries the argument

The central object is the recurrence network: each point of the time-delay-embedded phase-space trajectory is a node, and two nodes are linked when their distance falls below a chosen threshold $\varepsilon$, giving an adjacency matrix $A_{ij} = \Theta(\varepsilon - \|\vec{v}_i - \vec{v}_j\|) - \delta_{ij}$. The embedding uses dimension $m=4$ with time delay taken from the autocorrelation falling to $1/e$, and $\varepsilon=0.1$ is chosen as the value where most networks become just connected. The machinery works by translating geometric structure into network statistics: the small dense loop and larger ring of the cardiac attractor produce the two peaks in degree, and increasing $\varepsilon$ from 0.1 to 0.5 gradually connects the network, exposing two scaling regimes in link density whose log-log slopes define $\gamma_1$ and $\gamma_2$.

What would settle it

Take the 96 recordings and, for each, compute the degree distribution over a grid of thresholds from 0.05 to 0.5 and embedding dimensions from 2 to 6. If a disease class is bimodal at one threshold and unimodal at neighboring thresholds, the bimodality is not robust. Separately, fit the link-density-versus-threshold curve on log-log axes without pre-chosen breakpoints; if two distinct linear segments cannot be identified consistently, the claimed two scaling regions do not exist.

Watch

Extended reading notes

Core claim

The paper's central discovery is that recurrence networks built from short clinical ECG time series have a bimodal degree distribution, with two well-separated peaks in the probability of a node having a given number of links, and the authors state this is the first time bimodality has been reported for recurrence networks derived from time series. Underlying the claim is the reconstructed phase-space attractor: healthy ECGs show a small dense loop-like region and a larger ring, and the dense region produces the second peak at higher degrees. The paper further finds that link density $LD$ varies with the recurrence threshold $\varepsilon$ as $LD \sim \varepsilon^{\gamma}$ in two distinct scaling regions, with exponents $\gamma_1$ and $\gamma_2$ that differ across diagnostic classes. Disease-specific average values of clustering coefficient, average path length, link density, and the two scaling exponents are presented as quantifiable signatures, with healthy cases showing the least variability.

Load-bearing premise

The analysis assumes that one fixed embedding dimension and one fixed recurrence threshold are appropriate for every ECG class, so if that choice creates or hides the second peak in some classes, the bimodality is an artifact of parameter selection rather than a property of cardiac dynamics.

Editorial extensions

If this is right

Healthy ECGs show the least variability in clustering coefficient and average path length, so those measures provide a stable baseline for detecting abnormality.
Because the two scaling exponents differ across diagnostic classes, $\gamma_1$ and $\gamma_2$ can serve as compact quantitative features for separating cardiac conditions.
The method works on one-minute recordings, allowing complexity quantification for short and nonstationary clinical data where multifractal analysis would need longer series.
The bimodality separates ECG recurrence networks from networks of standard chaotic, hyperchaotic, and noise signals, placing cardiac dynamics in a distinct structural class.
In bundle branch block cases the second peak appears only when the threshold is increased, showing that the underlying structure exists but is less dense than in healthy cases.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the bimodality is robust, a simple classifier could use the positions and heights of the two degree peaks alone; the paper does not test classification accuracy, so this is an extension, not a claim.
The two scaling regimes might correspond to fast local recurrences within the dense loop and slower returns around the larger ring, a dynamical interpretation the paper does not explicitly make.
A direct test of the mechanism would be to generate synthetic ECG-like signals with and without the small-scale loop; if the second peak disappears when the loop is removed, the geometric interpretation would be confirmed.
Applying the same threshold-scanning analysis to heart-rate-variability series could connect this network signature to the larger HRV classification literature.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Two things to know. First, this paper reports something I haven't seen in the recurrence-network literature: bimodal degree distributions in RNs built from short ECG recordings, plus a two-regime scaling of link density with recurrence threshold. Second, the paper's own results contradict its headline. In Section 4 it says that for BB the distribution is unimodal at ε=0.1 and only becomes bimodal at 0.2, and for MI the first peak disappears as ε approaches 0.5. So "bimodality common to all types" is an overstatement.

What's actually new and useful: the observation itself is new relative to the cited RN reviews and multifractal ECG work, and the comparison with Rössler, Chen, and white noise gives a point of reference. The idea of varying ε and extracting scaling indices is a reasonable way to get more from short data. The authors are upfront about secondary diagnoses and about needing larger samples. They use public PTB data, standard preprocessing, and there is no circularity in the method.

Soft spots, in proportion. The central claim rests on visual inspection of Fig. 3; there is no quantitative bimodality test, no error bars on the distributions, and no surrogate comparison. That matters because the feature is threshold-dependent for at least one class. The scaling indices in Table 2 overlap heavily—healthy γ2 = 0.387 ± 0.296 vs MI γ2 = 0.372 ± 0.276—so disease-specific quantification is not actually demonstrated. The paper also acknowledges that for BB there is almost no slope change and for MI the regions are hard to discern. Those are honest statements, but they undercut the "two scaling regions" as a universal signature.

So the core findings are plausible and worth pursuing, but the claims as written are stronger than the evidence. This is a proof-of-concept, not a classification tool.

Who should read it: people working on network-based physiological time-series analysis who want a short-data alternative to multifractal methods. It deserves a serious referee. I'd recommend sending it to peer review with a clear request: replace visual inspection with quantitative bimodality tests, add surrogate data, fit the scaling regions with an explicit algorithm, and temper the "all types" language. If the authors do that, the observation could become a solid contribution. As it stands, I wouldn't cite it as evidence for disease-specific indices, but I'd cite it as an early report of a novel empirical feature.

Referee Report

4 major / 6 minor

Summary. The manuscript constructs recurrence networks (RNs) from one-minute clinical ECG recordings taken from the PTB database, for healthy subjects and four disease classes (Bundle Branch Block, Cardiomyopathy, Dysrhythmia, Myocardial Infarction), and compares them with RNs from chaotic, hyperchaotic, and noise data. The two central claims are, first, that the RN degree distribution of ECG data is bimodal, reflecting two spatial scales in the reconstructed attractor, and second, that the link density as a function of the recurrence threshold displays two scaling regions, quantified by indices gamma1 and gamma2. The authors further suggest that these features and their averaged network measures can distinguish disease classes from healthy cases. The analysis uses standard time-delay embedding with fixed embedding dimension m=4 and a fixed recurrence threshold epsilon=0.1, and reports averaged network measures in Table 1 and scaling indices in Table 2.

Significance. If the bimodality and two-regime scaling claims were rigorously established, this would be a useful contribution to the analysis of short clinical ECG recordings, since recurrence-network measures are relatively simple to compute and do not require long data sets. The manuscript has strengths: it applies a well-defined, standard recurrence-network construction to public data; it includes comparisons with standard dynamical systems and noise; and it reports average values over a reasonably large set of 96 recordings. However, the central claims are not currently supported by the evidence presented. The paper's own text states that at the chosen threshold epsilon=0.1 the degree distribution for Bundle Branch Block is unimodal, directly contradicting the abstract's claim that bimodality is a common feature of RNs from all types of ECG data. The scaling claim is similarly weakened by statements that for Myocardial Infarction the scaling regions are difficult to discern and for Bundle Branch Block there is almost no change of slope. No quantitative bimodality test, no threshold-invariance analysis, and no surrogate-data comparison are given.

major comments (4)

[Degree distribution and Abstract/Conclusion] The central claim that bimodality is a common feature of all types of ECG data is internally inconsistent with the reported results. In the section 'Degree distribution', the authors state for BB: 'the distribution is unimodal at this threshold', and in the threshold-variation section they state that for MI 'the first peak disappears as we increase epsilon to 0.5'. The abstract and conclusion nevertheless assert that both bimodality and two-region scaling are 'common features of RNs from all types of ECG data'. Since the fixed construction threshold epsilon=0.1 is the only threshold used for the headline degree distributions in Fig. 3, the universal claim is contradicted by the manuscript's own observations. This must be resolved, either by revising the claims or by providing a systematic threshold-dependent definition of bimodality that is tested across all 96 datasets.
[Construction of Recurrence Networks from embedded phase space attractors of ECG data] The choice of a single recurrence threshold epsilon=0.1 and a single embedding dimension m=4 for all classes is not justified for the central claims. The text justifies epsilon=0.1 only by saying 'most of the networks become just connected', which is a connectivity criterion, not a criterion for preserving bimodality or scaling behavior. The authors themselves note that BB is unimodal at epsilon=0.1 and becomes bimodal only at epsilon=0.2, while MI changes qualitatively as epsilon increases. Because the claimed two-scale attractor structure and the scaling indices depend on threshold, the manuscript needs to show that the qualitative features are stable over a meaningful range of epsilon and m, or to explain explicitly why a single fixed threshold is appropriate for the universal claims made in the abstract.
[Scaling of link density with recurrence threshold, Table 2] The two-scaling-region claim is not supported as a common feature. The text states that for MI 'it is difficult to discern the distinct scaling regions' and for BB 'there is almost no change of slope'. These statements contradict the general assertion that two scaling regions are observed for ECG data. In addition, the scaling indices in Table 2 show large standard deviations and heavy overlap across classes; for example, healthy gamma2 = 0.387 +/- 0.296 and MI gamma2 = 0.372 +/- 0.276 are statistically indistinguishable. The manuscript also does not state how the boundaries of the scaling regions are chosen. The authors need to provide a quantitative, reproducible procedure for identifying scaling regions, and should test whether gamma1 and gamma2 actually separate disease classes using a statistical test rather than visual inspection of Fig. 5.
[Degree distribution, Fig. 3] Fig. 3 shows degree distributions for single typical cases without error bars, uncertainty bands, or a quantitative bimodality test. The claim that the bimodal shape is a stable property of the recurrence-network ensemble requires at minimum a measure such as the dip test, Hartigan's dip test, or a Gaussian-mixture fit applied to all 96 datasets, together with a comparison against surrogate data or randomized thresholds. As presented, the evidence for bimodality is visual inspection of a few examples, which is not sufficient for a novelty claim of the strength made in the abstract.

minor comments (6)

[Abstract] There is a typo in the abstract: 'we also show that that there is relevant information' should read 'we also show that there is relevant information'.
[Fig. 3 caption and Degree distribution] The Fig. 3 caption states 'The degree distribution is bimodal in nature in all cases of ECG data sets', but the text in the same section says that BB is unimodal at the chosen threshold. The caption and the text should be brought into agreement.
[Construction of Recurrence Networks from embedded phase space attractors of ECG data] The downsampling procedure from 60000 to 6000 points by binning is described only in one sentence. The authors should specify the bin width, whether the binning uses averaging or another rule, and show that the downsampling does not remove the small-scale structures that are central to the bimodality interpretation.
[Scaling of link density with recurrence threshold] The manuscript states that epsilon is varied from 0.1 to 1.0 in steps of 0.01 but results are shown only up to 0.5 because 'there is no significant change after 0.5'. The criterion for 'no significant change' is not given, and the scaling regions in Fig. 5 appear to be selected by eye; a reproducible method for locating the two regions is needed.
[Conclusion] The conclusion says that the analysis 'is to be applied to larger number of data sets so that disease specific ranges of measures and scaling indices can be derived' and that this work is 'already in progress'. This statement is in tension with the earlier claims that disease-specific variations can be quantified from the present 96 datasets; the manuscript should clarify which claims are established now and which are prospective.
[General] No data availability or code availability statement is provided, although all data come from a public database. A statement with the exact PTB record identifiers and, where possible, the analysis code would substantially increase reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper reports empirical characterizations of ECG recurrence networks and does not fit a parameter into the quantity it claims to derive.

full rationale

Walking the claimed derivation chain: the recurrence network is built from standard definitions (adjacency from thresholded distances; CC, APL, LD from Eqs. 1-3), and all reported outputs -- degree distribution, clustering, path length, link density, and scaling indices -- are computed directly from the ECG data after fixed preprocessing. No equation defines a claimed output in terms of the conclusion it supports. The embedding dimension is chosen by False Nearest Neighbours, the time delay by the standard autocorrelation criterion, and the recurrence threshold is selected only by the 'just connected' condition ('most of the networks become just connected at epsilon = 0.1'), not by fitting to bimodality or to two scaling regions. The bimodality and two-scaling-region observations are empirical descriptions rather than predictions derived from the inputs; the later 'two spatial scales' explanation is an interpretation after the fact. The self-citations to the authors' prior work ([10], [18]) concern multifractal background and an embedding method and are not load-bearing for the central claim. The paper's own caveats -- e.g. that BB is unimodal at epsilon = 0.1 and that MI's first peak disappears at epsilon = 0.5 -- point to a possible threshold-dependence or internal-consistency problem, but that is a robustness concern, not circularity: the observed features are not equivalent to the construction parameters by definition. No fitted parameter is renamed as a prediction, and no uniqueness claim is imported from self-citation. The derivation chain is therefore self-contained; no circular step can be exhibited.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claims rest on standard recurrence-network methodology plus several dataset-specific choices: fixed epsilon and m, PTB label reliability, and the assumption that preprocessing retains relevant structure. No new entities are introduced. The scaling indices in Table 2 are fitted values, not independent predictions.

free parameters (4)

Recurrence threshold epsilon = 0.1
Chosen as the smallest threshold at which most networks become connected; degree distributions and scaling behavior are computed at this threshold, and bimodality is threshold-dependent (BB shows a second peak only at epsilon=0.2).
Embedding dimension m = 4
Set uniformly to 4 because FNN gave m=4 or less for most datasets; the visual bimodality and network measures depend on this choice.
Scaling indices gamma1 and gamma2 = See Table 2, e.g., healthy 0.782 +/- 0.288 and 0.387 +/- 0.296
Obtained from log-log fits to link density versus threshold in hand-chosen scaling regions; these are fitted characterizations, not predictions.
Downsampling factor = 10 (60000 to 6000 points)
Data are binned to 6000 points; the choice affects phase-space density and hence degree distribution but is not justified quantitatively.

assumptions (5)

domain assumption The ECG time series can be treated as a low-dimensional dynamical system suitable for time-delay embedding (Takens' theorem).
The paper reconstructs the attractor with m=4 and autocorrelation-based tau, implicitly assuming the ECG is generated by a low-dimensional deterministic process.
ad hoc to paper A single recurrence threshold epsilon=0.1 and embedding dimension m=4 are appropriate for all 96 datasets regardless of disease class.
Chosen for uniformity after observing that most networks become connected at epsilon=0.1, but disease-specific attractor widths suggest the optimal parameters may differ.
domain assumption The PTB database labels and the chosen 96 cases are sufficiently accurate and representative for disease comparisons.
The analysis compares healthy, BB, CM, DR, and MI groups, and the authors note secondary diagnoses can obscure results.
ad hoc to paper Filtering at 0.5 to 50 Hz, normalizing to [0,1], and downsampling by binning from 60000 to 6000 points preserve the features relevant to recurrence networks.
The paper states the process retains essential features but provides no quantitative check that the downsampled signal preserves the recurrence structure.
ad hoc to paper Link density scales as LD ~ epsilon^gamma over two regions, and the hand-chosen region boundaries are meaningful.
The two-scaling-region claim is supported only by visual inspection of log-log plots and fitted slopes, with no automated or statistical criterion for the regions.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Bimodality and Scaling in Recurrence Networks from ECG data." pith.science (2026). https://pith.science/paper/7BNN6L53

@misc{pith2026190801286,
  author       = {Pith},
  title        = {Pith review of: Bimodality and Scaling in Recurrence Networks from ECG data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7BNN6L53}},
  note         = {Machine review of arXiv:1908.01286}
}

read the original abstract

Human heart is a complex system that can be studied using its electrical activity recorded as Electrocardiogram (ECG). Any variations or anomalies in the ECG can indicate abnormalities in the cardiac dynamics. In this work, we present a detailed analysis of ECG data using the framework of recurrence network (RN). We show how the measures of the recurrence networks constructed from ECG data sets, can quantify the complexity and variability underlying the data. Our study shows for the first time that the RN from ECG show the unique feature of bimodality in their degree distribution. We relate this to the complex dynamics underlying the cardiac system, with structures at two spatial scales. We also show that that there is relevant information to be extracted from the scaling of measures with recurrence threshold. Thus we observe two scaling regions in the link density for ECG data which is compared with scaling in RNs from standard chaotic and hyperchaotic systems and noise. While both bimodality and scaling are common features of RNs from all types of ECG data, we find disease specific variations in them can be quantified.

Figures

Figures reproduced from arXiv: 1908.01286 by the authors.

**Figure 3.** Degree distributions in recurrence networks from [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

**Figure 5.** Variation of link density with recurrence threshold [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Variation in degree distribution with ε for different classes of ECG data sets. Scaling of link density with recurrence threshold. The changes in the structure of RN as ε is varied, can be studied using the link density. This variation as ε in the range 0.1 to 0.5 is …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 19 canonical work pages

[1]

and Schreiber T., Nonlinear time series anal- ysis Vol

Kantz H. and Schreiber T., Nonlinear time series anal- ysis Vol. 7 (Cambridge university press) 2004

work page 2004
[2]

V., Zou Y., Donges J

Donner R. V., Zou Y., Donges J. F., Marwan N. and Kurths J., New Journal of Physics , 12 (2010) 033025

work page 2010
[3]

V., Marwan N., Donges J

Zou Y., Donner R. V., Marwan N., Donges J. F. and Kurths J., Physics Reports, (2018)

work page 2018
[4]

Sharma V., The open cardiovascular medicine journal, 3 (2009) 110

work page 2009
[5]

R., Takamasu K., Stanley H

Kotani K., Struzik Z. R., Takamasu K., Stanley H. E. and Yamamoto Y., Physical Review E, 72 (2005) 041904

work page 2005
[6]

E., Clifford G

McSharry P. E., Clifford G. D., Tarassenko L. and Smith L. A. , IEEE transactions on biomedical engineer- ing, 50 (2003) 289

work page 2003
[7]

et al., Medical and Biological Engineer- ing and Computing , 42 (2004) 288

Acharya R., Kumar A., Bhat P., Lim C., Kannathal N., Krishnan S. et al., Medical and Biological Engineer- ing and Computing , 42 (2004) 288

work page 2004
[8]

R., Joseph K

Acharya U. R., Joseph K. P., Kannathal N., Lim C. M. and Suri J. S. , Medical and biological engineering and computing, 44 (2006) 1031

work page 2006

Show all 20 references

[9]

R., Fujita H., Oh S

Acharya U. R., Fujita H., Oh S. L., Hagiwara Y., Tan J. H. and Adam M. , Information Sciences , 415 (2017) 190

2017
[10]

M., Kotriwar Y., Harikrishnan K

Shekatkar S. M., Kotriwar Y., Harikrishnan K. and Ambika G. , Scientiﬁc reports, 7 (2017) 15127

2017
[11]

G., Taylor J

V alenza G., Citi L., Garcia R. G., Taylor J. N., Toschi N. and Barbieri R. , Scientiﬁc reports, 7 (2017) 42779

2017
[12]

and Schnabel A

Bousseljot R., Kreiseler D. and Schnabel A. , phy- sionet. org, (2004)

2004
[13]

L., Amaral L

Goldberger A. L., Amaral L. A. N., Glass L., Hausdorff J. M., Ivanov P. C., Mark R. G., Mietus J. E., Moody G. B., Peng C.- K. and Stanley H. E. , Circulation, 101 (2000 (June 13)) e215 circulation Electronic Pages: http://circ.ahajournals.org/content/101/23/e215.full PMID:108...

doi
[14]

and Kantz H., Chaos: An Interdisciplinary Journal of Nonlinear Science , 25 (2015) 097610

Bradley E. and Kantz H., Chaos: An Interdisciplinary Journal of Nonlinear Science , 25 (2015) 097610

2015
[15]

M., Spaan J

Acharya R., Krishnan S. M., Spaan J. A. and Suri J. S. , Advances in cardiac signal processing (Springer) 2007

2007
[16]

K., Uysal A

Berkaya S. K., Uysal A. K., Gunal E. S., Ergin S., Gunal S. and Gulmezoglu M. B. , Biomedical Signal Processing and Control, 43 (2018) 216

2018
[17]

and Schreiber T

Hegger R., Kantz H. and Schreiber T. , Chaos: An Interdisciplinary Journal of Nonlinear Science , 9 (1999) 413

1999
[18]

and Kemb- havi A

Harikrishnan K., Misra R., Ambika G. and Kemb- havi A. , Physica D: Nonlinear Phenomena , 215 (2006) 137

2006
[19]

, Networks: an introduction (Oxford univer- sity press) 2010

Newman M. , Networks: an introduction (Oxford univer- sity press) 2010

2010
[20]

and Yuan Z

Chen Z., Yang Y., Qi G. and Yuan Z. , Physics Letters A, 360 (2007) 696. p-6

2007

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [1]

and Schreiber T., Nonlinear time series anal- ysis Vol

Kantz H. and Schreiber T., Nonlinear time series anal- ysis Vol. 7 (Cambridge university press) 2004

work page 2004

[2] [2]

V., Zou Y., Donges J

Donner R. V., Zou Y., Donges J. F., Marwan N. and Kurths J., New Journal of Physics , 12 (2010) 033025

work page 2010

[3] [3]

V., Marwan N., Donges J

Zou Y., Donner R. V., Marwan N., Donges J. F. and Kurths J., Physics Reports, (2018)

work page 2018

[4] [4]

Sharma V., The open cardiovascular medicine journal, 3 (2009) 110

work page 2009

[5] [5]

R., Takamasu K., Stanley H

Kotani K., Struzik Z. R., Takamasu K., Stanley H. E. and Yamamoto Y., Physical Review E, 72 (2005) 041904

work page 2005

[6] [6]

E., Clifford G

McSharry P. E., Clifford G. D., Tarassenko L. and Smith L. A. , IEEE transactions on biomedical engineer- ing, 50 (2003) 289

work page 2003

[7] [7]

et al., Medical and Biological Engineer- ing and Computing , 42 (2004) 288

Acharya R., Kumar A., Bhat P., Lim C., Kannathal N., Krishnan S. et al., Medical and Biological Engineer- ing and Computing , 42 (2004) 288

work page 2004

[8] [8]

R., Joseph K

Acharya U. R., Joseph K. P., Kannathal N., Lim C. M. and Suri J. S. , Medical and biological engineering and computing, 44 (2006) 1031

work page 2006