Context-based near-lossless compression of EEG signals using neural network predictors

Abstract

Compression and transmission of electroencephalography (EEG) signal for telemedicine applications have made the stringent demand on the quality of the reconstructed signal. Although lossless compression scheme involving predictor–encoder satisfies the perfect reconstruction quality, the amount of compression it yields is very less thereby utilizing more bandwidth. Facing this reality, more and more users in medical community accept near-lossless methods to compress signal more effectively while maintaining a tight numerical bound on the maximum allowable error δ between the original and the reconstructed medical signals. Even though effective decorrelation is performed through the prediction step, there is sufficient structure remaining in the quantized residue that can be further explored by suitable error modeling schemes, thereby the effect of compression can be improved further. This paper discusses a context-based near-lossless compression using neural network predictors for the compression of EEG signals. Four neural network models, namely, single-layer and multi-layer perceptrons, Elman network and a generalized regression neural network are used and the results are compared with adaptive linear predictors such as LMS adaptive FIR filter and autoregressive (AR) model. It is found from the experimental results that the single-layer perceptron predictor yields the best results (at δ = 3 ) among the different predictors used with a saving of 2.5 and 0.51 bits when using lossless and near-lossless compression scheme without error modeling.