Benchmarking beat classification algorithms

The aim of this study is to compare the accuracy of a range of advanced and classical pattern recognition algorithms for beat and arrhythmia classification from ECG using a principled statistical framework. These are to be used in an application where no patient-specific adaptation of features or model is possible, which means that models must be able to generalise across subjects. Our results demonstrate that non-linear classification models offer significant advantages in ECG beat classification and that with a principled approach to feature selection, pre-processing, and model development, it is possible to get robust inter-subject generalisation even on ambulatory data.