Fusing QRS detection and robust interval estimation with a random forest to classify atrial fibrillation

This year's PhysioNet/CinC challenge aims to stimulate the development of robust algorithms to classify whether a short single-lead ECG recording shows normal sinus rhythm, atrial fibrillation (AF), an alternative rhythm, or is too noisy to be classified. Since the dataset consist of more than 8500 recordings, sophisticated methods from the realm of data fusion and machine learning can be applied. The approach presented here fuses timing information obtained via QRS detection with features from a robust interval estimator as well as waveform features using a Random Forest classifier. A super feature vector consisting of 78 global and 390 moving window features is proposed. Recursive feature elimination is used to select 25 features for the final algorithm. Using 10-fold cross-validation on the training dataset, the average scores F1n = 0.88, F1a = 0.77, F10 = 0.72, and F1 = 0.79 were achieved. On the full hidden test set, these values were 0.89, 0.78, 0.68, and 0.78 respectively.