The features extracted by de Chazal concentrate on the X, Y, and Z axes. They fall into three main categories:
Other features were ratios and differences between other features. Some of these features were extracted from the ECG literature, while others were developed within the Biomedical Systems Laboratory at the University of New South Wales. In total, 229 features were generated. In many cases, this includes both sines and cosines of angles to make presentation to the learning system simpler.
In de Chazal's work, a variety of learners were applied. The two main families examined were voted softmax neural networks and C5.0. He also employed a number of feature selection techniques.
The best accuracy value of 71.3 per cent (28.7 per cent error) was obtained by using 100 voted softmax-neural networks, each with different initial conditions applied to all features. These results compare extremely well with median values for human cardiologists of 70.3 per cent (a human panel obtained approximately 74.8 per cent accuracy).
However, there are some issues with de Chazal's work. Firstly, because neural networks were used (and in particular, 100 voted neural networks) the results are not very comprehensible. Clearly it would be desirable, especially from a medico-legal standpoint, for explanations for classifications of ECGs to be given. Secondly, de Chazal, from his work obviously a brilliant and dedicated PhD student, building on decades of research, spent at least 3 years developing software for extracting features. In other domains, such background knowledge may not be available.