Fuzzy C-Means Clustering and Sonification of HRV Features

Debanjan Borthakur; Harishchandra Dubey; Kunal Mankodiya; Paul Batchelor; Victoria Grace

REVIEW 4 major objections 5 minor 20 references

Fuzzy C-Means Clustering and Sonification of HRV Features

T0 review · 4 major / 5 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 Fuzzy C-means can separate three meditation groups from heart-rate-variability features, and the paper sonifies the average heartbeat interval as vowel-like sound.

desk verdict A clear, honest early-stage design note whose central cluster-separation claim rests on visual inspection of 12 hand-picked subjects; the idea is worth refereeing, but the evidence needs quantitative support. read the letter →

arxiv 1908.07107 v2 pith:JGAEH7WJ submitted 2019-08-19 cs.HC cs.LGcs.SDeess.ASstat.ML

Debanjan Borthakur , Victoria Grace , Paul Batchelor , Harishchandra Dubey , Kunal Mankodiya This is my paper · ORCID

classification cs.HCcs.LGcs.SDeess.ASstat.ML

keywords heartratevariabilityfuzzyC-meansclusteringsonificationmeditationbiofeedbackAVNNformantsynthesistime-domainHRVfeatures

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 tries to establish that fuzzy C-means clustering can identify which heart-rate-variability features best separate people practicing different meditation techniques, and that the average normal-to-normal interval (AVNN) is a good feature to sonify for biofeedback. Using data from Chi meditation, Kundalini Yoga, and spontaneous-breathing groups, the authors compute time-domain HRV features, cluster pairwise combinations with fuzzy C-means, and report that SDNN versus AVNN and RMSSD versus AVNN form three distinct clusters while pNN50 pairings do not. They take that separation as the basis for choosing AVNN, map it to sound through a formant-synthesis vocal sonification, and present spectrograms that differ across the three groups. The paper frames this as an early step toward a real-time, sound-based biofeedback training system, and it is explicit that the sonification itself has not yet been quantitatively evaluated.

What carries the argument

The two mechanisms that carry the argument are fuzzy C-means clustering and formant synthesis. Fuzzy C-means is a soft clustering method in which each data point carries a degree of membership in each of three clusters, found by minimizing the objective function $J_m = \sum_i \sum_j u_{ij}^m \|X_i - C_j\|^2$, where $u$ is the membership and $C$ the cluster center; the paper applies it pairwise to standardized HRV features: AVNN (average time between normal heartbeats), SDNN (standard deviation of those intervals), RMSSD (root mean square of successive differences), and pNN50 (percentage of successive intervals differing by more than 50 ms). The second mechanism is the sonification: the AVNN values are mapped to an audio signal in which a bandlimited narrow pulse wave passes through four Butterworth bandpass filters in series tuned to tenor-vowel formants, with the data controlling an $\alpha$ parameter that interpolates between two vowel states. That design is meant to make the heartbeat-derived signal more intelligible and memorable than a simple pitch mapping.

What would settle it

Run fuzzy C-means on all subjects in the public meditation dataset, not just the four chosen per group, using SDNN versus AVNN and RMSSD versus AVNN; if the three clusters no longer align with the Chi, Yoga, and normal-breathing labels, the central claim fails. A complementary check is a listening test in which naive participants classify sonified AVNN excerpts by meditation type; chance-level performance would undermine the claim that the sonification carries the cluster information.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that unsupervised fuzzy C-means clustering of z-scored time-domain HRV features separates the three meditation groups, with AVNN as the feature that makes the separation visible. The authors state that FCM clustering plots of SDNN versus AVNN and RMSSD versus AVNN show three distinct clusters accurately, whereas pairs involving pNN50 do not. Because the clusters are distinct, AVNN is selected as the sonification feature; the authors then construct a formant-synthesis sonification of AVNN whose spectrograms look different for Chi, Yoga, and normal breathing. The paper treats this as evidence that clustering can guide feature choice for sonification and, ultimately, that a listener could use sound to monitor their own meditation-related HRV state.

Load-bearing premise

The load-bearing premise is that the four subjects picked from each meditation group represent the group, so the three distinct clusters seen in the plots come from the meditation practices and not from who happened to be selected.

Editorial extensions

If this is right

If the cluster separation is real, AVNN, SDNN, and RMSSD can serve as features for classifying meditation type and for biofeedback that tells a user which meditative state their heart rhythm resembles.
The pairwise FCM procedure gives a visual, unsupervised way to screen HRV features before building a sonification, so future systems can pick the audible feature by data separation rather than by guesswork.
Because the goal is real-time feedback, the success criterion shifts from classification accuracy to learnability: a meditator should be able to recognize their sonified AVNN pattern and adjust breathing to change it.
The paper also implies that non-time-domain features, such as frequency-domain or nonlinear HRV measures, could be screened the same way to see whether they separate the groups even more clearly.

Reading between the lines

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

An implicit consequence the authors do not draw: the real test of the sonification is whether listeners can tell the meditation groups apart by ear. A forced-choice listening experiment using the sonified AVNN audio would settle that, and it is exactly the evaluation the paper defers.
The four-subject-per-group sampling means the cluster plots could be an artifact of subject choice; re-running the same clustering on the full public dataset would show whether the three-cluster separation generalizes. This is an extension of the paper's own procedure, not a claim the paper makes.
The same pairwise FCM screening could be applied to frequency-domain or nonlinear HRV features to see whether even cleaner separation exists; the paper restricts itself to time-domain features, so this remains an open test.
If the vocal sonification proves memorable in a controlled test, the mapping from AVNN to vowel space could be extended to other HRV features, creating a multidimensional auditory display rather than a single-feature voice.

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

Dear colleague,

This is a short, honest exploratory paper that combines two known techniques — fuzzy C-means clustering and sonification — in a specific new way: using FCM on time-domain HRV features to pick a feature for vocal formant-synthesis sonification. The combination is not in the cited literature, and the authors are upfront that it is early work. The writing is clear, the pipeline is reproducible in principle (Matlab toolbox, PhysioNet data), and the cluster-center table gives useful numerical context.

What is new is the particular mapping of AVNN to a tenor-vowel formant filter. That design choice is not evaluated in any rigorous way, and the paper says so. The evidence for the central claim — that FCM plots 'show three distinct clusters accurately' — is visual inspection of 2D plots from 12 subjects, four per group, selected without criteria. The authors themselves note that Euclidean distances between centers are not shown. A permutation test or cluster validity index would have addressed the most obvious objection, which is that fuzzy c-means with c=3 will always find three clusters, and 4 subjects per group can make the separation reflect age, breathing rate, or recording session rather than meditation style. The word 'accurately' is not supported.

The sonification evaluation is even softer: a non-blinded A-B test with four listeners, two of whom were musicians, and the tool being tested was developed by the authors' lab. The result — four out of four preferred the vocal synthesis — is anecdote, not data. There is also a citation slip: the SoX spectrogram step is credited to [12], the authors' own SmartEAR paper, when [19] is SoX's homepage. That should be fixed.

Still, the paper does not try to hide its weaknesses. It explicitly defers quantitative measures and a proper perceptual evaluation to future work. As an early-stage design note, it has some value: the cluster-center table and the spectrograms give a concrete starting point for a real biofeedback system. For a reader working on HRV sonification or mindfulness feedback, this is a useful sketch. I would not cite it for a strong claim, but I might mention it as an example of the feature-selection-by-clustering approach.

Recommendation: if this comes across your desk, send it to peer review, but the referee should insist on at least a cluster-validity check and a proper listener study before acceptance. The core idea is fine; the evidence is the problem.

Referee Report

4 major / 5 minor

Summary. This paper presents a pilot study that applies fuzzy C-means (FCM) clustering to time-domain heart rate variability (HRV) features extracted from three groups of subjects using data from PhysioNet: chi meditation, Kundalini yoga meditation, and spontaneous breathing. Four subjects per group (12 total) are selected. The authors cluster six pairwise feature combinations and report that plots of SDNN versus AVNN and RMSSD versus AVNN show three distinct clusters, leading them to select AVNN as the HRV feature to sonify. They describe a formant-synthesis sonification method and compare it informally with a simple pitch mapping via an A-B test with four listeners. The paper is explicitly framed as early steps toward a real-time sound-based biofeedback training system, and it defers quantitative cluster validation and formal sonification evaluation to future work.

Significance. If the cluster-separation claim were rigorously established, the paper would provide a useful data-driven rationale for choosing a single HRV feature for sonification in biofeedback systems, and the formant-synthesis mapping would be a plausible design candidate. The paper's strengths include its use of a public dataset, the presentation of cluster centers in Table I, and its transparent acknowledgment of missing quantitative validation. However, the central evidence is visual inspection of cluster plots from 12 subjects whose selection criteria are not stated, and the sonification evaluation is a non-blinded four-person A-B test. The contribution is therefore preliminary; the claims as stated exceed what the presented evidence supports.

major comments (4)

[Section II-C and Section III] The manuscript selects four subjects per group without stating the selection criteria or a representativeness argument (Section II-C), then asserts in Section III that FCM plots of SDNN versus AVNN and RMSSD versus AVNN 'shows three distinct clusters accurately.' With n=12 and the number of clusters fixed at three, FCM will always produce three cluster centers, so visual separation by itself does not establish that the clusters correspond to meditation type. The paper needs cluster-validity indices (e.g., silhouette coefficient, partition coefficient, or Dunn index) or a permutation test to show that the observed separation exceeds chance, and it must justify that the four chosen subjects per group are representative of the corresponding meditation practice.
[Section III] The authors explicitly state that 'Euclidean distance measurement of each pair of feature center points will provide more quantified assessment which is not shown in this work.' This admission is load-bearing because the choice of AVNN for sonification rests entirely on visual inspection of the cluster plots. Without quantitative separation measures or a comparison across feature pairs, the claim that SDNN versus AVNN and RMSSD versus AVNN separate the groups 'accurately' is not supported, and the subsequent feature selection for sonification is not justified on evidence beyond one author's visual reading.
[Section III, sonification evaluation] The only evaluation of the formant-synthesis sonification is an informal A-B test with four individuals (two musicians and two non-musicians) who reported that the vocal synthesis was 'more interesting, and easily memorable.' The test is non-blinded, reports no protocol or task definition, and includes no statistical or inter-rater analysis, so it cannot support claims about the sonification's comprehensibility, learnability, or effectiveness. The paper's own statement that 'no quantitative measures have been taken yet to evaluate the efficiency of this sonification technique' accurately describes the evidence: there is currently no empirical support for the claimed advantage of the vocal synthesis method.
[Section II-D and III] The same small dataset is used both to select AVNN as the sonification feature (via visual cluster inspection) and to demonstrate the sonification output. This is a selection-on-the-same-data issue: the chosen feature is not tested on independent data or on a held-out subset, so the sonification pipeline's performance is unknown. The paper should either validate the selected feature on a separate sample, apply cross-validation, or clearly frame the result as a hypothesis rather than a validated design choice.

minor comments (5)

[Abstract and Introduction] The word 'metrices' appears in the abstract and introduction; it should be 'metrics.'
[Section II-D] The phrase 'Fuzzy partition matrix overlap= 2.0' is nonstandard; the parameter is the fuzziness exponent m in the FCM objective function, and should be labeled as such.
[Section III, SoX citation] Reference [12] is cited for the SoX tool, but [12] is the SmartEAR paper on smartwatch-based unsupervised learning and is unrelated to SoX; the SoX manual or website should be cited instead.
[Table I] The column header 'STDNN' is likely a typo for 'SDNN.'
[Section III] There are typographical errors: 'We adpot' should be 'We adopt,' and 'WAV file' is rendered as 'W A V file.'

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the clustering step is descriptive, and the only self-citation is background material.

full rationale

The paper makes no derivation or predictive claim that reduces by construction to its inputs. Its central result is an empirical observation that FCM plots of SDNN vs AVNN and RMSSD vs AVNN show visually distinct clusters for the three meditation groups, and AVNN is then chosen for sonification on that basis. This is a data-exploration and feature-selection workflow, not a fitted parameter being relabeled as a prediction: the same small dataset is used both to generate the plots and to choose AVNN, which is an evidentiary weakness (confirmation on the same sample, no validation or permutation test), but it is not circular because the claim 'these features separate the groups' is not entailed by the definition of the features or the clustering algorithm. The self-citation [12] is used only to describe K-means and to reference a spectrogram-rendering tool; it is not load-bearing for the paper's conclusions. The authors even explicitly state that a more quantified Euclidean-distance assessment 'is not shown in this work,' acknowledging the absence of a formal validation rather than constructing one from prior assumptions. Any concerns about representativeness of the 12-subject sample or visual-only cluster assessment are correctness and evidence risks, not circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No new entities are introduced. The free parameters are standard FCM settings chosen by hand. The main assumptions concern the correctness of the feature extraction, the representativeness of the small sample, and the untested link between visual cluster separation and sonification interpretability.

free parameters (3)

Number of FCM clusters (Nc) = 3
Set to match the three known meditation groups; not determined by data-driven criteria.
Fuzziness exponent (m) = 2.0
Chosen as the Matlab fcm default and a common choice; not tuned against external benchmarks.
Maximum FCM iterations = 100
Convergence setting chosen by hand; not reported as sensitive.

assumptions (4)

domain assumption The mhrv toolbox computes HRV features correctly from RR intervals.
Invoked in Section II-D as the sole feature extraction method; no validation on this dataset is provided.
domain assumption The four subjects chosen from each PhysioNet group are representative of Chi meditation, Kundalini Yoga, and spontaneous breathing.
Loaded in Section II-C; the paper does not state selection criteria, and the cluster plots rely entirely on these 12 subjects.
domain assumption Visual separation in 2D pairwise FCM plots is a valid indicator of feature usefulness for sonification.
Used in Section III to select AVNN; no listening test or classification experiment connects cluster separation to sonification interpretability.
standard math FCM objective function converges to a useful local minimum for the chosen parameters.
Assumed by using Matlab's fcm; this is a standard property but not verified for this data.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Fuzzy C-Means Clustering and Sonification of HRV Features." pith.science (2026). https://pith.science/paper/JGAEH7WJ

@misc{pith2026190807107,
  author       = {Pith},
  title        = {Pith review of: Fuzzy C-Means Clustering and Sonification of HRV Features},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JGAEH7WJ}},
  note         = {Machine review of arXiv:1908.07107}
}

read the original abstract

Linear and non-linear measures of heart rate variability (HRV) are widely investigated as non-invasive indicators of health. Stress has a profound impact on heart rate, and different meditation techniques have been found to modulate heartbeat rhythm. This paper aims to explore the process of identifying appropriate metrices from HRV analysis for sonification. Sonification is a type of auditory display involving the process of mapping data to acoustic parameters. This work explores the use of auditory display in aiding the analysis of HRV leveraged by unsupervised machine learning techniques. Unsupervised clustering helps select the appropriate features to improve the sonification interpretability. Vocal synthesis sonification techniques are employed to increase comprehension and learnability of the processed data displayed through sound. These analyses are early steps in building a real-time sound-based biofeedback training system.

Figures

Figures reproduced from arXiv: 1908.07107 by the authors.

**Figure 2.** Fuzzy -C Means Clustering respectively. The results are z score normalized prior to clustering by Fuzzy C-Means algorithm. We have used Matlab’s fcm function for clustering: [centers,U] = fcm(data,Nc) , for clustering the heart rate variability metrices that are calculated using the mhrv toolbox. Here centers is the center of the clusters, U is the fuzzy partition matrix. Nc is the number of clusters [PITH_FULL_IM… view at source ↗

**Figure 5.** Spectrogram of the sonified features of Meditation Techniques Yoga [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Spectrogram of the sonified features of Normal breathing [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 20 canonical work pages

[12]

(2018, September)

Borthakur, D., Peltier, A., Dubey, H., Gyllinsky, J., & Mankodiya, K. (2018, September). SmartEAR: Smartwatch-based Unsupervised Learning for Multi-modal Signal Analysis in Opportunistic Sensing Framework. In 2018 IEEE/ACM International Conference on Connected Health: Applications, Systems and Engineering Technologies (CHASE) (pp. 75-80). IEEE

work page 2018
[19]

Sound eXchange: HomePage. (n.d.). Retrieved August 3, 2019, from http://sox.sourceforge.net/l

work page 2019
[1]

George E. Billman. 2011. Heart rate variability - A historical per- spective. Frontiers in Physiology 2 NOV , November (2011), 113. https://doi.org/10.3389/fphys.2011.00086

work page Pith review arXiv 2011
[2]

Bhattacharya, D.N

Anilesh Dey, D.K. Bhattacharya, D.N. Tibarewala, Nilanjan Dey, Amira S. Ashour, and Dac-Nhuong Le. 2016. Chinese-chi and Kundalini yoga meditations effects on the autonomic nervous system: comparative study

work page 2016
[3]

Weippert, M., Behrens, K., Rieger, A., Kumar, M., & Behrens, M. (2015). Effects of breathing patterns and light exercise on linear and nonlinear heart rate variability. Applied Physiology,Nutrition, and Metabolism, 40(8), 762-768

work page 2015
[4]

A., Tavares, B

JPorto, A. A., Tavares, B. S., Vidigal, G., Garner, D. M., Raimundo, R. D., de Abreu, L. C.,...& Valenti, V . E. (2018). Nonlinear Dynamics of Heart Rate During Slow Breathing Exercise. Indian J Physiol Pharmacol, 62(2), 160-169. International Journal of Interactive Multimedia and Artiﬁcial Intelligence 3, 7 (2016), 87

work page 2018
[5]

Bernardi, L., Sleight, P., Bandinelli, G., Cencetti, S., Fattorini, L., Wdowczyc-Szulc, J., & Lagi, A. (2001). Effect of rosary prayer and yoga mantras on autonomic cardiovascular rhythms: comparative study. Bmj, 323(7327), 1446-1449

work page 2001
[6]

Yu, B., Feijs, L., Funk, M., & Hu, J. (2015). Designing auditory display of heart rate variability in biofeedback context. Georgia Institute of Technology

work page 2015

Show all 20 references

[7]

K., Bratec, S

Doll, A., Hlzel, B. K., Bratec, S. M., Boucard, C. C., Xie, X., Wohlschlger, A. M., & Sorg, C. (2016). Mindful attention to breath regulates emotions via increased amygdalaprefrontal cortex connectivity. Neuroimage, 134, 305-313

2016
[8]

(2019, April)

Bahameish, M. (2019, April). Can Changes in Heart Rate Variability Represented in Sound be Identiﬁed by Non-Medical Experts?. In Ex- tended Abstracts of the 2019 CHI Conference on Human Factors in Computing Systems (p. SRC01). ACM

2019
[9]

Dubus, G., & Bresin, R. (2013). A systematic review of mapping strategies for the soniﬁcation of physical quantities. PloS one, 8(12), e82491

2013
[10]

Blanco, A., Lorena, A., Grautoff, S., & Hermann, T. (2018). Car- dioSounds: A portable system to sonify ECG rhythm disturbances in real-time. Georgia Institute of Technology

2018
[11]

C., Glass, L., & Goldberger, A

Ballora, M., Pennycook, B., Ivanov, P. C., Glass, L., & Goldberger, A. L. (2004). Heart rate soniﬁcation: A new approach to medical diagnosis. Leonardo, 37(1), 41-46

2004
[13]

S., Murugappan, M., & Yaacob, S

Zheng, B. S., Murugappan, M., & Yaacob, S. (2013, April). FCM clustering of emotional stress using ECG features. In 2013 International Conference on Communication and Signal Processing (pp. 305-309). IEEE

2013
[14]

Exaggerated Heart Rate Oscillations During Two Meditation Techniques

Peng C-K, Mietus JE, Liu Y , Khalsa G, Douglas PS, Benson H, Gold- berger AL. Exaggerated Heart Rate Oscillations During Two Meditation Techniques. International Journal of Cardiology 70:101-107, 1999

1999
[15]

A., Rosenberg A

Behar J. A., Rosenberg A. A. et al. (2018) PhysioZoo: a novel open access platform for heart rate variability analysis of mammalian elec- trocardiographic data. Frontiers in Physiology

2018
[16]

R: A language and environment forstatistical computing

R Core Team (2017). R: A language and environment forstatistical computing. R Foundation for Statistical Computing, Vienna, Austria. URL https://www.R-project.org/

2017
[17]

(2006, January)

Jovanov, E. (2006, January). On spectral analysis of heart rate vari- ability during very slow yogic breathing. In 2005 IEEE Engineering in Medicine and Biology 27th Annual Conference (pp. 2467-2470). IEEE

2006
[18]

Csounds.com. (2019). Appendix D. Formant Values. [online] Available at: http://csounds.com/manual/html/MiscFormants.html

2019
[20]

K., Pradhan, C., Barik, R

Sahoo, A. K., Pradhan, C., Barik, R. K., & Dubey, H. (2019). DeepReco: Deep Learning Based Health Recommender System Using Collaborative Filtering. Computation, 7(2), 25

2019

Pith tools

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

[1] [12]

(2018, September)

Borthakur, D., Peltier, A., Dubey, H., Gyllinsky, J., & Mankodiya, K. (2018, September). SmartEAR: Smartwatch-based Unsupervised Learning for Multi-modal Signal Analysis in Opportunistic Sensing Framework. In 2018 IEEE/ACM International Conference on Connected Health: Applications, Systems and Engineering Technologies (CHASE) (pp. 75-80). IEEE

work page 2018

[2] [19]

Sound eXchange: HomePage. (n.d.). Retrieved August 3, 2019, from http://sox.sourceforge.net/l

work page 2019

[3] [1]

George E. Billman. 2011. Heart rate variability - A historical per- spective. Frontiers in Physiology 2 NOV , November (2011), 113. https://doi.org/10.3389/fphys.2011.00086

work page Pith review arXiv 2011

[4] [2]

Bhattacharya, D.N

Anilesh Dey, D.K. Bhattacharya, D.N. Tibarewala, Nilanjan Dey, Amira S. Ashour, and Dac-Nhuong Le. 2016. Chinese-chi and Kundalini yoga meditations effects on the autonomic nervous system: comparative study

work page 2016

[5] [3]

Weippert, M., Behrens, K., Rieger, A., Kumar, M., & Behrens, M. (2015). Effects of breathing patterns and light exercise on linear and nonlinear heart rate variability. Applied Physiology,Nutrition, and Metabolism, 40(8), 762-768

work page 2015

[6] [4]

A., Tavares, B

JPorto, A. A., Tavares, B. S., Vidigal, G., Garner, D. M., Raimundo, R. D., de Abreu, L. C.,...& Valenti, V . E. (2018). Nonlinear Dynamics of Heart Rate During Slow Breathing Exercise. Indian J Physiol Pharmacol, 62(2), 160-169. International Journal of Interactive Multimedia and Artiﬁcial Intelligence 3, 7 (2016), 87

work page 2018

[7] [5]

Bernardi, L., Sleight, P., Bandinelli, G., Cencetti, S., Fattorini, L., Wdowczyc-Szulc, J., & Lagi, A. (2001). Effect of rosary prayer and yoga mantras on autonomic cardiovascular rhythms: comparative study. Bmj, 323(7327), 1446-1449

work page 2001

[8] [6]

Yu, B., Feijs, L., Funk, M., & Hu, J. (2015). Designing auditory display of heart rate variability in biofeedback context. Georgia Institute of Technology

work page 2015