ECG classification based on time and frequency domain features using random forests

We present a combined method of classical signal analysis and machine learning algorithms for the automated classification of 1-lead ECG recordings, which was developed in the course of the Computing in Cardiology Challenge 2017. To classify ECG recordings into the four classes as defined for the Challenge (normal, suspicious to AF, suspicious to other arrhythmia, noise) we used MATLAB- and a set of algorithms for detection of beats, wave point detection on detected beats, quality evaluation of the detection, averaging of beats, beat classification, rhythm classification and many more. We extracted a variety of features from both time and frequency domain etc. as input features for the classifier. A total of 380 features were used to train a Random Forest-based classifier (bagged decision trees). Since classes for the Challenge were severely unbalanced, weights based on the class distribution were applied. To train the classifier and for our internal evaluation we used cross-validation on all available ECGS from the training-set. 10-fold cross-validated F1 score on the training set is 0.83. Final F1 score from the official challenge evaluation on the enhanced dataset is 0.81, which is quite close to the other top performing algorithms.