just out of curiosity, what kind of algorithm should I use if I want predict if an Electrocardiogram is normal or not ?
I've got a set of normal and pathological ECGs and I would like to train an automaton(?) to decide whether a new ECG is normal or not;
Thanks,
EDIT: I don't want to analyze the signal of the ECG: I would already have the coordinates X/Y of the main points of the QRS complex, but I want to find if there is a significant difference between the curves.
Also take a look at techniques already used for chromatogram classfication, as many of the issues were already adressed there (dynamic time warping, etc.)
Well this might come as a surprise, but I asked cardiologists the same question a few years ago. Like: "how can you give people a recorder that records an ECG for 24 hours and then see where an individual arythmic event did occur, what software do you use?" They said: "actually we use hardware, our eyes and a wall." So what they do is: they just hang the printout along the wall and then stand back far. That way your eyes, especially after some training, will be able to spot the irregularity. They were actually very satisfied about how well that worked, although they liked some help with hanging the printouts on the wall.
I think for this specific problem they might be right. It will be hard to beat the human eye in recognizing an irregularity that does not necessarily have a predefined direction (like longer S-T interval or something like that).
That's what my colleagues are doing here at the hospital: the physicians 'just' look at the ECGs and decide whether there is a irregularity. I was wondering if there was a tool/algorithm to mind those ECGS.
You could compare the raw ECG traces using [windowed] correlation in time or frequency domain. But, if you have QRS data, why not just use this reduced-dimension representation?
This is on e approach among others. Allows deriving a distance (althoug this is not strictly speaking a distance with all properties, but a good proxy).
Don"t know how easy it is ti find a QRS and suspect it's better ti restrict the method to the expected QRS region.
Anyways, other measures are currently used ;
Fuzzification of the Analysis of Heart Rate Variability Using ECG in Time, Frequency and Statistical Domains Ismail Sadiq, Shoab Ahmad Khan
ICCEA '10 Proceedings of the 2010 Second International Conference on Computer Engineering and Applications - Volume 01
Can't reach this paper and no guarantee but from the abstract, seems a good place to start with ...
I can't help you directly here. But I know that in the nineties of the last century this was a hot topic in Medical Informatics/Medical Physics. I know that there is actually a medical informatics company that initiated because of some successes with ecg prediction. But I really don't remember the name.
If I recall correctly it has to do with decomposing the signal into basic signals, but this is coming from the dungeons of my memory.
I do remember that back then it were mainly physicist active in the area.
Also take a look at techniques already used for chromatogram classfication, as many of the issues were already adressed there (dynamic time warping, etc.)
Thanks Khader. I already had your 1st ref. Developing a classifier is what I'm looking for.