Wavelet feature space in computer-aided electroretinogram evaluation

The paper discusses creating a wavelet-based feature space for the classification of transient pattern electroretinograms (PERGs)-signals utilized in ophthalmology to evaluate the state of the retina. Discrete wavelet transform (DWT) can provide compact signal description, which is more accurate than time-domain data. A procedure for the proper choice of transform parameters is proposed. Both time-domain and wavelet features of these waveforms are visualized using principal components analysis. Separability of feature spaces is compared using k-means clustering algorithm. The results suggest that PERG waveforms are better separable when represented by DWT coefficients of full time-domain signal, than in traditional peak-based feature space.