2.1. Dataset

In this study, data for signal quality assessment were collected from four sources. They are listed in the following.

Physionet/CinC Challenge (CinCHS) 2016 HS Database [19, 20]: these recordings were collected from various positions on the chest surface at different environments including home, hospital, and uncontrolled surroundings. It consists of 3153 recordings collected from 765 subjects. The detail description is given in [19].

Pascal Classifying Heart Sound Challenge (PASCAL) Database [21]: the data were collected from two sources. One was from an iPhone app, and another was from a clinic trial in a hospital using a digital stethoscope. There are 859 recordings available.

Heart Sounds Catania (CTHS) Database 2011 [22, 23]: this database was a collection of heart sounds used for biometry by the University of Catania, Italy. It contained heart sounds acquired from 206 people using a digital stethoscope. There are 412 recordings available. The data can be downloaded at [22].

Cardiac disease heart sound (CDHS) Database: It included 3875 recordings acquired by the authors' group from 76 patients in the second attached hospital of Dalian Medical University since 2015.

The sampling frequencies in the four datasets are different. They are 2000 Hz, 11025 Hz, 44100 Hz, and 2000 Hz in CinCHS, CTHS, PASCAL, and CDHS, respectively. The four databases provide 8299 recordings available. However, to ensure that the signal quality can be reliably assessed, those recordings with time length less than 6 s are excluded. It is found that the noise to cause low signal quality is mainly respiratory sounds, environment noise, and skin contact.