The paper presents a number of results regarding the construction of specific overcomplete dictionaries for ECG compressed sensing (CS). The dictionaries were built using normal and patological cardiac patterns extracted from 24 recordings of the MIT-BIH Arrhythmia Database. It has been shown that the compression results obtained using the CS concept based on specific dictionaries are better that those using the wavelet overcomplete dictionaries. Starting from the concept of sparse signal with respect to a given overcomplete dictionary the paper present several results regarding the possibility of simple pattern classification as well

In this paper we investigate the reconstruction of compressed sensed ECG signals using dictionaries specific for each patient. Two projection matrices, random and Bernoulli, have been tested. For 24 records from the MIH-BIH database and for a compression ratio (CR) of 10:1 an average percentage rootmean-square difference (PRD) of 0.81, a normalized PRD (PRDN) of 15.49 and a quality score (QS) of 12.34 have been obtained. For record 117, for CR = 15:1 the PRD, PRDN and QS were 0.96, 17.28 and 15.62 respectively.

This paper presents a robust reconstruction technique of electrocardiograph (ECG) signals in a compressed sensing based acquisition system, using custom complete and overcomplete dictionaries composed of real ECG patterns. Both signals and atoms are preprocessed segments of ECG recordings. We tested three types of projection matrices and found that the best reconstruction results are obtained when the projection matrix is the product of a random matrix with the transpose of the dictionary used for the sparse representation. Further improvements are obtained by reconstructing every signal multiple times from the same measurements, using different randomized dictionaries, and then averaging all the reconstructions.

Abstract-We present a robust method for the detection of the first and second heart sounds (s1 and s2), without ECG reference, based on a music beat tracking algorithm. An intermediate representation of the input signal is first calculated by using an onset detection function based on complex spectral difference. A music beat tracking algorithm is then used to determine the location of the first heart sound. The beat tracker works in two steps, it first calculates the beat period and then finds the temporal beat alignment. Once the first sound is detected, inverse Gaussian weights are applied to the onset function on the detected positions and the algorithm is run again to find the second heart sound. At the last step s1 and s2 labels are attributed to the detected sounds. The algorithm was evaluated in terms of location accuracy as well as sensitivity and specificity and the results showed good results even in the presence of murmurs or noisy signals.

No abstract