Electrocardiographic (ECG) signals can be used to analyze and even predict patients’ epileptic attacks. However, this requires real-time processing and ECG signal classification. Using probabilistic neural networks (PNNs), a small set of salient ECG features is sufficient for an accurate classification of epileptic seizures.

Übeyli has published various closely related studies, which diminishes the novelty of this paper. It starts with a concise overview of the signal-to-noise ratio (SNR) saliency measure, discrete wavelet transform (DWT), and PNN. Subsequently, the results on a standard dataset are presented: 98.33 percent accuracy, using only two features extracted by a DWT.

Regrettably, the paper contains some flaws. It handles a two-class problem (for example, an epileptic seizure or not) and no multi-class problem. Hence, why a PNN is chosen as a classifier is not clear. Moreover, an experimental study like this should compare its processing scheme with a few related schemes. Finally, the results section lacks content; except for two accuracy percentages, hardly any information is provided.

Although this paper has its limitations, it does pose a clear statement: the success of automatic classification depends on feature selection. Feature selection is also crucial in bringing offline pattern recognition to (online) real-time adaptable systems--for example, real-time detection of epileptic attacks. This is a message that cannot be repeated enough, since it is still too often forgotten.