qrs detector improvements

[cx1111]

This will be the single thread for discussing improvements for qrs detectors. I will probably only be working on the xqrs class in the near future.

Current thoughts:

• The filter used to accentuate the qrs (we'll call this the moving wave integration filter) is currently an upright ricker with a fixed width: wavelet_filter = signal.ricker(self.qrs_width, 4). There are many things we could improve upon, such as:
  • Inverting it if we figure out the r wave is inverted
  • Designing a set of template beats for different ecg channels.

[krivenkoz]

If you don't mind, i add several suggestions:

1. XQRS works pretty well and has good specificity, PPV and FPR for MIT-BIH Arrhythmia Database, The QT Database and others. But sometimes for real signals there are errors for R-peak localization. This error equals +/- one sample around of peak sample.
2. Adding some marks for erroneous/noisy beats (GQPOST functionality: changing N annotations into artifact ("|") annotations if they are likely to be erroneous)
3. Adding R-peak voltages in returning value of xqrs_detect()
4. Designing a set of template for different type of the beats. Like in CONSTRUE framework (Atrial fibrillation, Bigeminy, Trigeminy, Ventricular flutter/fibrillation, etc.). Link to framework below:
   [https://github.com/citiususc/construe]
5. Adding ECGPUWAVE functionality (P,T wave peak/limit locator). I don't know if XQRS is able to do it

[peiyaoli]

One possible solution is to call the raw C function through Cython. I used this approach in some of our project

[artur-deluca]

A bit off-topic but, is there any publication that describes the xqrs algorithm?

[Lucas-Mc]

Hey @artur-deluca

You can find a review of some algorithms from this link:

https://physionet.org/content/challenge-2014/1.0.0/

And more technical detail here:

http://ecg.mit.edu/george/publications/qrs-cinc-2003.pdf

However, checking out the source code documentation from these three links will also help:

wfdb-python/wfdb/processing/qrs.py

Lines 13 to 67 in 733f3e4

	class XQRS(object):
	"""
	The QRS detector class for the XQRS algorithm. The `XQRS.Conf`
	class is the configuration class that stores initial parameters
	for the detection. The `XQRS.detect` method runs the detection algorithm.
	The process works as follows:
	- Load the signal and configuration parameters.
	- Bandpass filter the signal between 5 and 20 Hz, to get the
	filtered signal.
	- Apply moving wave integration (MWI) with a Ricker
	(Mexican hat) wavelet onto the filtered signal, and save the
	square of the integrated signal.
	- Conduct learning if specified, to initialize running
	parameters of noise and QRS amplitudes, the QRS detection
	threshold, and recent R-R intervals. If learning is unspecified
	or fails, use default parameters. See the docstring for the
	`_learn_init_params` method of this class for details.
	- Run the main detection. Iterate through the local maxima of
	the MWI signal. For each local maxima:
	- Check if it is a QRS complex. To be classified as a QRS,
	it must come after the refractory period, cross the QRS
	detection threshold, and not be classified as a T-wave
	if it comes close enough to the previous QRS. If
	successfully classified, update running detection
	threshold and heart rate parameters.
	- If not a QRS, classify it as a noise peak and update
	running parameters.
	- Before continuing to the next local maxima, if no QRS
	was detected within 1.66 times the recent R-R interval,
	perform backsearch QRS detection. This checks previous
	peaks using a lower QRS detection threshold.
	Attributes
	----------
	sig : 1d ndarray
	The input ECG signal to apply the QRS detection on.
	fs : int, float
	The sampling frequency of the input signal.
	conf : XQRS.Conf object, optional
	The configuration object specifying signal configuration
	parameters. See the docstring of the XQRS.Conf class.
	Examples
	--------
	>>> import wfdb
	>>> from wfdb import processing
	>>> sig, fields = wfdb.rdsamp('sample-data/100', channels=[0])
	>>> xqrs = processing.XQRS(sig=sig[:,0], fs=fields['fs'])
	>>> xqrs.detect()
	>>> wfdb.plot_items(signal=sig, ann_samp=[xqrs.qrs_inds])

https://wfdb.readthedocs.io/en/latest/processing.html
https://www.physionet.org/physiotools/wag/gqrs-1.htm
https://www.physionet.org/physiotools/wag/wqrs-1.htm

Hope this helps!

[artur-deluca]

Thanks @Lucas-Mc!