Re: Sample Entropy

Hi Pamphile,
 
Did I understand you right that you don't think it is a good fit for SciPy?
 
Also, I did not know about antropy until now; thanks for pointing it out. Digging into the code, I found that it only uses numba for short sequences (<5k items) and delegates longer ones to MNE-Features (which I will have to check out). To see how well this works I ran a quick benchmark using random data. For short sequences, the runtime is mine < numba < MNE (yay), and for large sequences, the runtime is MNE < mine < numba. In other words, my kernel appears to be an improvement over the current numba one, so I will open an issue at antropy in any case.
 
Best,
Sebastian
 

> Pamphile Roy <[email protected]> hat am 17.01.2023 13:43 CET geschrieben:
>  
>  
>  
> Hi Sebastian,
>  
> This looks interesting.  I am not sure where that could/would fit in our API though.
>  
> Did you have a look at https://raphaelvallat.com/antropy https://raphaelvallat.com/antropy/build/html/index.html
> (from Raphaël the author of Pingouin)? It seems fast (uses Numba) and with quite some features.
>  
> Cheers,
> Pamphile
> 
> 
> > On 17 Jan 2023, at 12:48, Sebastian Wallkötter <[email protected]> wrote:
> > 
> > 
> 
> > Hello SciPy,
> >  
> > I implemented an algorithm to compute sample entropy (a time-series statistic) [0,1] and wonder if this is a useful addition to SciPy.
> >  
> > In a nutshell, sample entropy quantifies the amount of regularity of a series; for example, the sequence "0101010101..." has (relatively) low sample entropy whereas "011100101100..." has (relatively) high sample entropy. It is a modification of approximate entropy with less bias. Sample Entropy and approximate entropy are useful whenever we want to measure the amount of "turbulence" in a series, e.g. when quantifying the irregularity of a heart rate signal or hormone levels [2], or the turmoil of a stock's value [3]. There is also a two-variable version called cross (sample/approximate) entropy which measures the association between two non-stationary time series.
> >  
> > While conceptually simple a naive implementation of sample entropy doesn't scale well for large input sequences, as it operates on all pairs of fixed-length windows into the sequence. Hence why I think it is useful to have an optimized implementation available. The poor performance can be seen in the attached benchmark (blue line). Clever tricks on reusing intermediate results make this a little better (orange line), but only make sense if we avoid numpy, because they require frequent access to single elements which is expensive when done from python. Doing the same via a C extension (green line), however, makes the metric much more viable on larger sequences.
> >  
> > If this sounds interesting, I will create an enhancement issue for this and we can discuss the details there. If not, then I would be grateful if somebody could point me to another OSS repo that might find this useful.
> >  
> > Best,
> > Sebastian
> >  
> >  
> > [0] https://en.wikipedia.org/wiki/Sample_entropy
> > [1] https://journals.physiology.org/doi/epdf/10.1152/ajpheart.2000.278.6.H2039
> > [2] https://pubmed.ncbi.nlm.nih.gov/11797860/
> > [3] Olbryś, Joanna, and Elżbieta Majewska. "Regularity in stock market indices within turbulence periods: The sample entropy approach." Entropy 24.7 (2022): 921.
> > [result.png]_______________________________________________
> > SciPy-Dev mailing list -- [email protected]
> > To unsubscribe send an email to [email protected]
> > https://mail.python.org/mailman3/lists/scipy-dev.python.org/
> > Member address: [email protected]
> > 
> _______________________________________________
> SciPy-Dev mailing list -- [email protected]
> To unsubscribe send an email to [email protected]
> https://mail.python.org/mailman3/lists/scipy-dev.python.org/
> Member address: [email protected]
>

_______________________________________________
SciPy-Dev mailing list -- [email protected]
To unsubscribe send an email to [email protected]
https://mail.python.org/mailman3/lists/scipy-dev.python.org/
Member address: [email protected]