Skip to main content

Table 3 Deep learning-based analysis of respiratory sounds

From: The coming era of a new auscultation system for analyzing respiratory sounds

Topic of study

Number of subject/recording

Number of classes

Applied model

Result

References

Recognition of pulmonary diseases from lung sounds using CNN and LSTM

213/1483

6: Normal, asthma, pneumonia, bronchiectasis, COPD, heart failure

CNN, biLSTM

Accuracy

[12]

   

 biLSTM: 98.16%

 
   

 CNN: 99.04%

 
   

 CNN + biLSTM: 99.62%

 

Classification of respiratory sounds using OST and deep residual networks

Not available/489

3: Crackle, wheeze, normal

OST, ResNets

Accuracy

[13]

   

 STFT: 93.98%

 
   

 ST: 97.79%

 
   

 OST + ResNets: 98.79%

 

Detection of respiratory sounds based on wavelet coefficients and machine learning

130/705

3: Crackles, rhonchi, normal

SVM, ANN, KNN

Accuracy

[14]

   

 SVM: 69.50%

 
   

 ANN: 85.43%

 
   

 KNN: 68.51%

 

Benchmarking of eight recurrent neural network variants for breath phase and adventitious sound detection

279/9765

6: Inhalation, exhalation, wheeze, stridor, rhonchus, crackles

LSTM, GRU, BiLSTM, BiGRUc, CNN-LSTM, CNN-GRU, CNN-BiLSTM, and CNN-BiGRU

F1 score

[15]

   

 LSTM: 73.9%

GRU: 77.6%

 
   

 BiLSTM: 76.2%

BiGRU: 78.4%

 
   

 CNN-LSTM: 78.1%

CNN-GRU: 80.6%

 
   

 CNN-BiLSTM: 80.3%

CNN-BiGRU: 80.6%

 

Classification of lung sounds through DS-CNN models with fused STFT and MFCC features

Not available/12,691

4: Normal, wheeze, crackle, unknown

DS-CNN, VGG-16, AlexNet, DS-AlexNet, LSTM, GRU, TCNd

Accuracy

[16]

   

 DS-CNN: 85.74%

VGG-16: 85.66%

 
   

 AlexNet: 79.92%

DS-AlexNet: 80.86%

 
   

 LSTM: 76.92%

GRU: 78.50%

 
   

 TCN: 75.51%

 

Implementation of AI algorithms in pulmonary auscultation examination

50/522

4: Wheezes, rhonchi, fine and coarse crackles

Neural network

F1-score (%)

[17]

   

 Coarse crackles: Doctors (42.8%), NN (47.1%)

 
   

 Fine crackles: Doctors (51.1%), NN (64.6%)

 
   

 Wheezes: Doctors (61.8%), NN (66.4%)

 
   

 Rhonchi: Doctors (61.0%), NN (72%)

 

AI accuracy in detecting pathological breath sounds in children

25/192

2: Crackles, wheeze

Neural network

PPAa/NPAb

[18]

   

 Crackles: Clinicloud 0.95/0.99, Littman 0.82/0.96

 
   

 Wheezes: Clinicloud 0.90/0.97, Littman 0.80/0.95

 

Classification of lung sounds using CNN

1630/17,930

â‘  2: Healthy, pathological

CNN, SVM

Accuracy (CNN/SVM)

[19]

  

â‘¡ 3: Rale, rhonchus, normal

 

â‘  86%/86%

â‘¡ 76%/75%

 

Feature extraction technique for pulmonary sound analysis based on EMD

30/120

3: Normal, wheezes, crackles

ANN, SVM, GMM

Accuracy

[20]

   

 EMD with ANN: 94.16%

 
   

 EMD with SVM: 93.75%

 
   

 EMD with GMM: 88.16%

 

Application of deep Learning to classify the severity of COPD

Not available/120

2: Crackles, wheeze

NN, DBN

Accuracy

[21]

   

 NN: 77.60%

 
   

 DBN: 95.84%

 

Application of deep Learning to detect early COPD

50/600

1: Wheeze

DBN

Accuracy

[22]

   

 DBN: 93.67%

 

Application of semi-supervised deep learning to lung sound analysis

284/11627

2: Crackles, wheeze

SVM

 AUC

[23]

   

 Crackle: 0.74

 
   

 Wheeze: 0.86

 
  1. aPPA positive percent agreement
  2. bNPA negative percent agreement
  3. cBiGRU bidirectional gated recurrent unit
  4. dTCN temporal convolutional network, DBN deep belief networks, COPD chronic obstructive pulmonary disease