tam-pham / hrvstructure

Supplementary Materials for the study "Unveiling the Structure of Heart Rate Variability (HRV) Indices: A Data-driven Meta-clustering Approach"

Unveiling the Structure of Heart Rate Variability (HRV) Indices: A Data-driven Meta-clustering Approach

Abstract

Heart Rate Variability (HRV) can be estimated using a myriad of mathematical indices, but the lack of systematic comparison between these indices renders the interpretation and evaluation of results tedious. In this study, we assessed the relationship between 57 HRV metrics collected from 302 human recordings using a variety of structure-analysis algorithms. We then applied a meta-clustering approach that combines their results to obtain a robust and reliable view of the observed relationships. We found that HRV metrics can be clustered into 3 groups, representing the distribution-related features, harmony-related features and frequency/complexity features. From there, we described and discussed their associations, and derived recommendations on which indices to prioritize for parsimonious, yet comprehensive HRV-related data analysis and reporting.

Supplementary Materials

Reproducible analysis scripts as well as open-access data files can be found in this repository.

• Preprint
• Supplementary Materials

hrvstructure's Issues

Optimal complexity parameters

A useful and easy to do side study would be to, on the same data base, compute optimal complexity parameters (delay, embedding & threshold) on the raw heart heart rate and maybe analyze as a function of mean heart rate.

Which PSD method was used to calculate frequency based HRV metrics?

Hi Tam, interesting take to understand the forest of HRV metrics better. Thanks for open sourcing the data and the nuerokit2 package.

I have few questions.

Which PSD method was used to calculate frequency based HRV metrics? I guess you are using https://github.com/neuropsychology/NeuroKit/blob/master/neurokit2/hrv/hrv_frequency.py to calculate frequency based HRV metrics. Knowing the PSD method would be needed to reproduce your paper since the frequency metrics are dependent on PSD estimation methods.

And, what is the rationale behind picking the PSD method of your choice (over others)? Also is there a particular reason Welch method is the default method in https://github.com/neuropsychology/NeuroKit/blob/master/neurokit2/hrv/hrv_frequency.py

Did you perform any outliers detection step to remove outlier beats before calculating HRV metrics?

Roadmap

So after OHBM and meeting some people from the community, we should really push this forward.

I am tagging @danibene who expressed interest in helping out here and @sangfrois might find useful stuff as well.

Roadmap for paper

• Find and add more data (I would look openNeuro, also, best would be to gain access to big banks like UK biobank but the application process looks complicated. Francois also mentioned neuromod from which we could get some PPG data?)
• Add new HRV indices that we added since
  • Try the possibility of a new metric for the frequency domain analog to the 1/f slope used in EEG (foof package) to see how it compares to LF/HF
• add UMAP as part of the dimension reduction if we don't have it already (and maybe make a nice plot out of it as it's a trendy method)

Side projects / side papers (?)

• Have this set of databases used easily re-usable with a good documentation on how to download it (will be of interest to the physiopy community)
• A useful and easy to do side study would be to, on the same data base, compute optimal complexity parameters (delay, embedding & threshold) on the raw heart heart rate and maybe analyze as a function of mean heart rate.
• Use this metadatabase for quality control validation (?) @sangfrois

We could schedule a meeting to refresh ourselves and the new potential collaborators about what we've done so far, and lay out the work ahead