*3.1. The Test Signal Set*

The proposed adaptive ECG sampling algorithm was tested with the use of CSE Multilead Database [32] recommended for electrocardiographs performance tests by the IEC [26]. It is worth a remark that the CSE Database is used for industrial validation of ECG wave delineation accuracy, and the perceptual model is adapted to the information on wave start and endpoint positions. Consequently, tests with the CSE files are sufficient for a complete evaluation of the method and the algorithm performance in the case of ectopic or missing beats, arrhythmias, etc., can be inferred from the reliability of wave border detection in those cases.

We used 125 automatically annotated 12-lead signals from CSE Dataset 3. The Dataset 3 consists of proportionally represented examples of normal ECGs, myocardial infarction, bundle branch blocks, premature ventricular contractions, ischemic ST changes, atrial fibrillation and many others. It is sampled with a 12-bit resolution at 500 samples per second to a data stream of 72,000 bits per second (bps). The average additional data stream carrying the local sampling rate is 1034 bps (i.e., 1.44% of the original record's data volume).

We first compared adaptive ECG sampling results for representative heartbeats of each CSE record with use of either their reference or the Ascard 6-calculated wave border values. Although the Ascard 6 software (see Section 2.3) is certified, the respective fiducial points differ by up to 2–4 ms depending on the point type and the presence of noise. Substituting these values was possible with just writing a few lines of I/O instructions to the development version of the Ascard 6 software. With any of the 123 CSE files (pacemaker-originated records 67 and 70 were discarded for the lack of annotation) we did not notice any difference of sampling results. Consequently, we stated that the inaccuracy of wave delineation procedure within Ascard 6 software is not significant and can be neglected. In all remaining tests, additionally taking into account beat-to-beat variability in each 12 s ECG sequence, individual wave borderlines were determined with the Ascard 6 software.
