CODEWITHSUNDEEP
genetic-algorithmevolutionary-algorithm
Aman
Aman

Reputation: 153

Genetic algorithm approach to frequency detection (in a photoplethysomogram)

A photoplethysomogram is shown as below.

Example Photoplethysomogram

Basically you can see the brightness of the image rise and fall as a function of time. This output could be from a pulse oximeter which measures the blood flow of your finger.

My question is this, are there genetic algorithm (or generally: evolutionary computational) based approaches to figure out the frequency of the "drops". I know FFT (well DFT) could compute the frequency (or atleast give us a frequency domain representation of the input shown above).

If you had to solve this problem using Genetic Algorithm techniques, how would you approach it? (I am not looking for actual solutions, just your ideas on how the representation and fitness functions would look in a GA design).

Upvotes: 2

Views: 249

Answers (2)

tcbelding
tcbelding

Reputation: 46

This can be thought of as a special case of a more general problem, predicting dynamical systems (timeseries), and there's actually been quite a lot of work on applying genetic algorithms to that. E.g. see the discussion in http://www.amazon.com/Introduction-Genetic-Algorithms-Complex-Adaptive/dp/0262631857 pp. 56-61, or one of the original papers, by Norman Packard (one of the founders of chaos theory): http://www.ccsr.uiuc.edu/web/Techreports/1988-89/CCSR-89-10.pdf

-Ted

Upvotes: 1

Andreas
Andreas

Reputation: 6465

I don't think a GA is best suited to solve this, but since you asked specifically about a GA solution here are some thoughts. To me it seems the saddle points are better suited to be found since they are much sharper and you would get the frequency out of them as well.

I assume the problem data would be a double vector where each position holds the brightness at a certain time. I would require that all points in this vector are sampled from equidistant time slots. Then finding the frequency could be to find an offset and an interval which minimizes (or maximizes) the average of the values obtained from the problem data at the points given by offset+x*interval. The advatange of using the average is that you don't need to given a minimal interval since the average will drop once the interval becomes too low. Unfortunately, it could find an interval too high so you will also need to maximize the number of points at which the fitness is evaluated. This creates a tougher multi-objective problem.

Upvotes: 2

Related Questions