Lossless EEG signal compression

Efficient Sequential Compression of Multichannel Biomedical Signals

“…Preliminary results of this work were presented at the 2014 European Signal Processing Conference (EUSIPCO 2014), Lisbon, Portugal, 2014 Authors are with Universidad de la República, Montevideo, Uruguay (e-mails: {icapurro, fefo, almartin, nacho, exavier}@fing.edu.uy, and gseroussi@ieee.org) removing temporal and/or spatial correlations. This produces a predicted signal, which is then subtracted from the original signal to obtain a prediction error that is encoded losslessly [1], [2], [3], [4], [5]. Among prediction methods, typical choices include linear predictors [1], [3], and neural networks [2].…”

“…In relation to the objectives posed for this work, we note that algorithms such as those described in [6], [4], [7], [8] require a number of operations that is superlinear in the number of channels. Some methods, including [6], [3], [4], [7], [8], [10] also have the drawback of having to perform more than one pass over the input data. Arithmetic coders such as those used in [3], [4], [7], [8] are computationally more expensive than, for example, a Golomb coder [12], which is extremely simple and thus popular in low-power embedded systems.…”

“…This produces a predicted signal, which is then subtracted from the original signal to obtain a prediction error that is encoded losslessly [1], [2], [3], [4], [5]. Among prediction methods, typical choices include linear predictors [1], [3], and neural networks [2]. After correlation is (hopefully) removed by the prediction stage, residuals are encoded according to some statistical model.…”

“…Some methods, including [6], [3], [4], [7], [8], [10] also have the drawback of having to perform more than one pass over the input data. Arithmetic coders such as those used in [3], [4], [7], [8] are computationally more expensive than, for example, a Golomb coder [12], which is extremely simple and thus popular in low-power embedded systems. Finally, methods such as [4], [7], [8] split the signal into an approximate transform and a residual, and perform an exhaustive search among several candidate splittings, where each candidate splitting is tentatively encoded, thus further increasing the coding time.…”

“…The database consists of 1308 2-minute recordings of subjects performing a motor imagery task (DB1a), and 218 1-minute calibration recordings (DB1b). • DB2a and DB2b [31] (BCI Competition III, 3 The database consists of 373 recordings with an average duration of 3.5 minutes, a minimum of 3.3 minutes, and a maximum of 5.5 minutes. • DB5 [29], [34] (Physikalisch-Technische Bundesanstalt (PTB) Diagnostic ECG Database): standard 12-lead, 1000Hz, 16bps ECG.…”