You will need to develop your own QRS detector / delimitator. From there you will measure whatever features you find important. The obvious ones are RR interval QT segment, TQ segment, amplitude of QRS, width of QRS and area. There are people working for years on this. It is not a trivial issue.

Dear colleague,

You entered a Terra Incognita. There are many possible ways how to proceed. Some of them follow:

a) You detect QRS complexes and manage to detect missing P-waves.

b) You can use the software on the page https://physionet.org to detect QRS complexes.

c) You can scan literature how to extract features from ECGs and use them to discriminate curves.

d) Possibly Matlab has a function detecting QRS efficiently. I am not using Matlab.

The field is wide and broad, it takes time to orientate yourself. One name could help you to start effectively: @Alberto_Porta (https://www.researchgate.net/profile/Alberto_Porta).

Good luck :-)

Dear Amirali Mirsajadi ,

You can use convolutional neural networks (CNN), which learns both features and classifier simultaneously. In the context of human activity recognition based on wearable sensor data (i.e., ECG), many works have presented state-of-the-art results using CNN.

In my GitHub page (https://github.com/arturjordao/WearableSensorData), you can find implementations and results in the context of wearable sensor data. In addition, on the page, there are papers which proposes features and classifiers.

I hope this to be useful for you.

As other researchers have already pointed out, the task you would like to perform is not trivial. A plenty of studies about PVC detection have been published, and several methods have been proposed, from those based on the simple analysis of consecutive RR interval duration values (Cuesta et al., 2014) to more sophisticated approaches, such as Bayesian classification models (Casas et al., 2018). Depending on the strategy you would like to apply, certain preprocessing steps may be required, such as noise filtering (inspect signal quality to understand what you need) and EGG fiducial point detection (at least, the QRS complex). Concerning the classification stage, it is hard to give a unique answer. Each of the models you mentioned have pros and cons. Certainly, your choice will be influenced by the size of your database and the a priori knowledge about the signals. If your databse is not very big and you have identified some relevant ECG features, you may try with SVM and other classical approaches. If the dataset is big enough, you may try with neural networks and try to classify PVC vs non*PVC beats. You can provide the beat themselves as input, if you have no clue about the features to extract Hope it is useful, good luck!