Table 2 Common respiratory/lung sound datasets in the literature
From: A review on lung disease recognition by acoustic signal analysis with deep learning networks
Dataset name | Description | Used by | Source |
|---|---|---|---|
Respiratory Sounds Dataset (RSD) [68] | Regular sound signals in addition to three kinds of adventitious respiratory sound signals: wheezes, crackles, and a combination between wheezes and crackles | [68] | |
HF_Lung_V1 | Comprises 9765 lung sound audio files (each lasting 15 s), 18,349 exhalation labels, 34,095 inhalation labels, 15,600 irregular adventitious sounds’ classes, and 13,883 regular adventitious sound classes (including, 4740 rhonchus classes, 8458 wheeze classes, and 686 stridor classes) | [66] | [66] |
Respiratory-database@TR | Each patient has 12-channel lung sounds. Short-term recordings, multi-channel analysis, 5 COPD (chronic obstructive lung disease) severity levels (COPD4, COPD3, COPD2, COPD1, COPD0) (At least 17Â s). This dataset was considered by | [13] | [12] |
Own generated database | The lung sounds were captured using an e-stethoscope and an amplifier linked to a laptop. An e-stethoscope with a chest piece that is touched by the patient and a microphone-based recording sound signals with a 44,100 Hz sampling rate that is attached to signal amplifiers are used in this setup. The amplifier kits extend the signal range to about (70–2000 Hz) with respiratory sounds (with frequency controller and control amplifier) when associated with an earphone (to listen to live records) and a PC | [21] | [21] |
Own generated database | Data is separated into two types: Sub-interval set, which includes complete patient set, which comprises all patients' measures and is classed as Abnormal or Normal, counting all patients' sub-interval measurements of any duration. It has around 255Â h of measured lung sound signals | [46] | [46] |
Own generated database | RSs non-stationary data collection with 28 separate patient records. For training and testing, two distinct sets of signals were employed. Except for crackles and wheezes, which were data from six patients each, each class in the training and test sets comprised two recordings from distinct patients. The sampling frequency of the recorded data was 44.1Â kHz | [122] | [122] |
R.A.L.E. repository | It is a collection of digital recordings of respiratory sounds in health and sickness. These are the breath sounds that physicians, nurses, respiratory therapists, and physical therapists hear using a stethoscope when they auscultate a patient's chest. Try-R.A.L.E. Lung Sounds, which provides a vast collection of sound recordings and case presentations, as well as a quiz for self-assessment | [22] | [159] |
R.A.L.E. lung sounds 3.0 | It includes five regular breathing recordings, four crackling recordings, and four wheeze recordings. To eliminate DC components, a first-order Butterworth high-pass filter with a cut-off frequency of 7.5Â Hz was employed, followed by an eighth-order Butterworth low-pass filter with a cut-off frequency of 2.5Â kHz to band restrict the signal | [4] | [123] |
Respiratory sound database | It developed by two Portuguese and Greek research teams. It has 920 recordings. The duration of each recording varies. 126 patients were recorded, and each tape is documented. Annotations include the start and finish timings of each respiratory cycle, as well as if the cycle comprises wheeze and/or crackle. Wheezes and crackles are known as adventitious noises, and their presence is utilized by doctors to diagnose respiratory disorders | [134] |