Medical information, especially in time-varying forms, typically creates large databases and requires the relevant large storage media and strategies. Different applications may impose additional requirements: for example, machine-learning techniques may well operate off-line and thus alleviate the need for real time availability, while emergency processing for early detection needs as much data as possible on demand.
For transmission, the choice of medium is another critical factor, since the difference in bandwidth capacity between T3 networks and normal telephony networks (PSTN) is on the order of 1000 to 1. Data compression may be an obvious option here, but this also presents a number of interesting questions, especially when lossy compression techniques are put to use, thereby removing possibly critical information from the data set. In addition to this, the huge amounts of data produced by the acquisition equipment may also be a serious obstacle.
The compression of digital information comes in two forms: lossless, where data is encoded reversibly and, as a result, decoding will faithfully reproduce the initial data set; and lossy, in which facets of the data set are sacrificed to achieve bigger gains in file size and bandwidth for transmission. There are several apparent advantages to either approach. However, the most important of each side are the trade-off between compression reversibility and complexity on the one hand and the resulting size of information and required bandwidth on the other.
In the case of lossless encoding, typical compression rates are in the order of 2 to 1, meaning that the size of the file produced by the compression algorithm is half of the initial file. However, all information contained in the original data set is retained and can be retrieved with minimal decoding effort.
When lossy algorithms are employed, however, a portion of the data is deemed either unnecessary to retain or almost correlated with another portion and thus can be dismissed.
Besides the lossy/lossless division, one may categorize compression efforts with respect to their analogy to signal representation and processing. For example, one can distinguish between algorithms operating on temporal representations and the frequency domain. Usually the particular technique that is used depends on the actual signal to be compressed. In the case of electrocardiograms (ECGs), a compression ratio of 7.8 has been reported using differential pulse code modulation (DPCM), an algorithm that uses linear prediction in order to de-correlate the samples of the input signal. The rationale behind this is that data samples with higher probabilities contribute less to what is already known and, since they occur more often than the rest, are assigned to smaller codewords. This means simply that they are represented with less bits in a look-up table. Reversely, sample values that do not arise as frequently, and have small probabilities, can be assigned to larger representations, a fact that does not harm file size since the values in question are infrequent.
This lossless process, which is called Huffman encoding, can be performed in both 1D and 2D signals, making it a very important tool for data compression. It has to be noted though that the remaining modules of the DPCM compression algorithm, and especially quantization of the values of the error sequence that is produced by the prediction/decorrelation step, introduce irreversible loss of information with a higher compression rate.
DPCM is a typical example of algorithms used to prepare the signal for either storage or transmission; signal encoding may also be performed to segment the signal or detect prominent features from it. In this framework the Amplitude Zone Time Epoch Coding (AZTEC) algorithm can be used to assist in analyzing ECGs, producing QRS detection automatically. This is another example of irreversible loss of information that permits higher compression rates.
One may describe this method as adaptive downsampling, where temporally static or almost static information within a specific time window is replaced by a constant value. To reconstruct the waveform, linear interpolation between the remaining samples is sufficient. However, this produces a lower resolution waveform that has an inherently jagged look.
