RespSounds

random reading notes

Dataset	Year	Country	#Sam. (#Sub.)	Sounds	Device	Respiratory conditions	Annotation
ICBHI	2017	Greece, Portugal	6898 (126)	Lung sounds	Stethoscope， Microphone	Cycle-level: crackle, wheeze; subject-level: COPD, LRTI, URTI	Expert-label
Respiratory Database @TR	2020	Turkey	504 (42)	lung sounds	Stethoscope 12 channel	COPD 0-4	Expert-label
Stethoscope (KAUH)	2021	Jordan	336 (112)	Lung sounds	Stethoscope	Cycle-level: inhalation, exhalation, crackle, wheeze; subject-level: asthma, COPD, BRON, pneumonia, heart failure, lung fibrosis	Expert-label
HF Lung V1 [V2]	2021 [2022]	Taiwan	9765 (279) [13957 (300)]	Lung sounds	Stethoscope	Cycle-level: inhalation, exhalation, wheeze, stridor, rhonchus, DAS; subject-level: acute respiratory failure, COPD, pneumonia, and so on	Expert-label
SPRSound	2022	China	9089 (292)	Lung sounds	Stethoscope	Cycle-level: Normal, Rhonchi, Wheeze, Stridor, Coarse Crackle, Fine Crackle, and Wheeze & Crackle; record-level: CAS, DAS, poor quality; subject-level: asthma, bronchitis, pneumonia (non-severe), pneumonia (severe), and other	Expert-label
Pulmonary Sound	2022	India	532 (unknown)	Lung sounds	Stethoscope	Normal/sick; Crepitation (C), Normal (N), Rhonchi (R) and Wheezing (W)	Expert-label

Dataset	Year	#Sam. (#Sub.)	Sounds	Device	Respiratory conditions	Annotation
Pertussis	2016	38 (38)	Cough	Microphone	Pertussis, asthma, croup, BRON	Self-report
Pfizer	2018	6593 (unknown)	Audio	Microphone	contains diseased sounds (coughing and sneezing)	BMAT
ESC-50 + FADKaggle2018	2021	40+273	Cough		cough
EPFL multimodal	2023	1440 (15)	Audio w noise (cough, breath, laugh, throat clearing)	2 Microphone + IMU	4300 Cough events	semi-automatic

lung sounds

• Normal lung sound spans in the frequency range 100– 1000 Hz and is devoid of any discrete peaks
• Adventitious sounds
  • Wheezes: continuous sound having a musical character, characterized by periodic waveforms with a dominant frequency usually over 100 Hz and with a duration of >= 100 ms, common sign of obstructive lung disease
  • Crackles: discontinuous, explosive and transient; less than 20 ms, and frequency content typically is wide
  • Stridors: very loud wheezes, consequence of a morphologic or dynamic obstruction in larynx or trachea, characterized by a prominent peak at about 1000 Hz.
  • Squawks: short inspiratory wheezes that occur primarily in restrictive lung diseases; always occur along with crackles, often begin with a crackle; duration rarely exceeds 400 ms; assumed to originate from oscillation of small airways after sudden opening, and their timing seems to depend on the transpulmonary pressure in a similar manner as in crackles.
  • Rhonchi: often have a low-pitched, rattling, rumbling or bubbling quality; may even sound similar to wheezes on occasion; may have an even more liquid sound than either wheezes or crackles, but could also sound dry; dominant frequency of less than 200 Hz.
  • examples https://hawaiicopd.org/lung-sounds/

Kandaswamy, 2004

• The stethoscope has a frequency response that attenuates frequency components of the lung sound signal above about 120 Hz and the human ear is not very sensitive to the lower frequency band that remains.
• normalize by loudness before any analysis

Sengupta, 2016

• 5 cepstral-based statistical features, including LFCC and MFCC

Bardou, 2018

• CNN outperforms traditional ML with features
• CNN structure

Grønnesby

• Feature Extraction for Crackle Detection

cough sounds

Jaclyn A Smith, 2006

• Healthcare professionals are actually poor at making diagnoses base on cough sounds; they are more used to breathing sounds.
• Cough is the commonest symptom for which patients seek medical advice.
• spectrogram showing and illustrating interruptions, wheezes, etc

Feature Extraction for the Differentiation of Dry and Wet Cough Sounds (a master thesis)

• Common characteristics of both dry and wet cough signals were explained
• The first feature extraction algorithm computes the number of peaks of the energy envelope of the cough signals and the second feature extraction algorithm extracts the power ratio of the two frequency bands of the second phase of the cough signals.

Use of Cough Sounds for Diagnosis and Screening of Pulmonary Disease，2017

• 54 patients + 33 HC, voluntary cough

• also collected: auscultation, clinical questionnaire, and peak flow meter, full pulmonary function test (gold standard)

• a general framework for cough sound analysis( automatic cough segmentation + feature extraction + classification design)

  • three evidence-based features (variance, kurtosis, and zero crossing irregularity) + an additional feature developed (rate of decay)

• cough sounds surprisingly had better performance than lung sound auscultation alone, but had significantly lower performance compared to our clinical questionnaire or peak flow meter test

• moderate performance, comparable or better to lung sounds, however not as good as (and cannot improve) questionnaire or flow meter

Detection of covid-19 from joint time and frequency analysis of speech, breathing and cough audio ICASSP 2022

• fusion three modalities, using time-frequency features (spectrogram, mfcc)
• ablation on three modalities as well as dropping frequency domain

Covid-19 Sounds

Brown et al., KDD20

• handcrafted feature + VGGish, of breathing and coughing, with LR/SVM classifier
• different combination of modalities for different tasks

Xia et al., NeurIPS 2021, datasets and benchmarks

• fine-tuned VGGish best, 3modality/cough best

Dang et al., 2022, longitudinal disease progression prediction

• GRU to predict current status and disease trajectory, no future forcasting

Dang et al., Interspeech 2022, semi-supervised learning Fixmatch

• Supervised training for labelled samples + weakly augmented spectrograms output as pseudo label for strongly augmented spectrograms

Dang et al., KDD 2023, Conditional Neural ODE for progression forecasting

• forecast multiple steps in the future
• encoder (fit distribution and then sample, using fc) before latent ODE and decoder (fc) after
• outperforming RNN and transformer