Biomedical Signal Data Features Dimension Reduction Using Linear Discriminant Analysis and Threshold Classifier in Case of Two Multidimensional Classes

The discriminant linear analysis algorithm is a statistical tool, without iterative training, capable of determining the best data projection reducing the characteristics dimension size to a new space. The new space contains a dimension number equals to the number of input classes minus one. In the specific case of two input classes, the resulting space projection will be a one-dimensional space where a simple threshold classifier is able to determine that a sample belongs to one or another class. This paper explains the theory and also this special case, applying the technique on biomedical data like pulmonary crackles characterization, and detection of spikes in EEG signals. Besides machine learning techniques, statistical analysis proved to be a simple, fast and efficient way to classify patterns.