Related works | Sample size | Purpose | Features | Performance | Issues |
---|---|---|---|---|---|
Aykanat et al. [67] | 40 IPF, 211 COPD, and 574 other pulmonary single or mixed conditions, 805 healthy subjects | Binary classification (healthy vs pathological) and 12-class lung disease classification | SVM, k-NN, GB | Performance for binary classification: Acc 88% to 92% Se 85% to 92% Sp 85% to 88% Performance for lung disease classification: Acc 43% to 68% Se 53% to 96% | Severity classification not done. LUS not used |
Charleston-Villalobos et al. [68] | 19 ILD (12 IPF, 7 extrinsic allergic alveolitis), 8 healthy subjects | Binary classification (healthy subjects vs patients with ILD) | AR model, SNN | Performance evaluation of the neural network: Acc 76.9% to 98.8%, Se 80.4% to 100%, Sp 73.3% to 100% | Severity classification not done. LUS not used |
Flietstra et al. [69] | 39 IPF, 95 CHF, 123 PN | Binary classification (IPF vs CHF and IPF vs PN) | BPNN, SVM | Acoustic properties of fine crackles of IPF help distinguish them from crackles of CHF and PN | No healthy subjects. Severity classification not done. LUS not used |
Fukumitsu et al. [70] | 71 ILD (24 honeycombing + , 47 honeycombing-) | Predicting honeycombing on HRCT by the acoustic properties of fine crackles | FFT | Acoustic properties of fine crackles distinguish the honeycombing from the non-honeycombing group | No healthy subjects. Severity classification not done. LUS not used |
Horimasu et al. [71] | 34 ILD, 8 COPD or asthma, 7 lung tumor, 5 lung nodule, 6 other | Comparisons of machine-learning-based quantification of four types of lung sounds between lung fields with and without ILD in HRCT and chest X-ray | Described in [72] | AUROC 0.855, Acc 75%, Se 76.1%, Sp 73.6% | Severity classification not done. Results suffered from the presence of background noise. LUS not used |
Kahya et al. [73] | 23 ILD, 28 COPD, 18 healthy | Binary classification (healthy vs pathological) | AR model, k-NN | Acc 71.1% | Severity classification not done. LUS not used |
Kim et al. [74] | 112 ILD, 211 COPD, 497 other pulmonary conditions, 51 healthy | Binary classification (healthy vs pathological) and three-class (crackles vs wheezes vs rhonchi) classification | CNN | Performance for binary classification: AUROC 0.93 Acc 86.5% Performance for abnormal sounds classification: AUROC 0.92 Acc 85.7% | Severity classification not done. LUS not used |
Malmberg et al. [75] | 8 fibrosing alveolitis, 8 emphysema, 8 asthma, 8 healthy subjects | Diagnosis agreement between clinical and machine-learning-based classification of lung sounds | FFT, SOM | Kappa for fibrosing alveolitis 0.54 | Small number of patients. Severity classification not done. LUS not used |
Manfredi et al. [76] | 98 CTD patients (42 ILD + , 56 ILD-) | Identifying CTD patients with possible ILD using lung sound-based binary classification (presence vs absence of ILD) compared with chest HRCT | FFT | Acc 82.6%, Se 88.1%, Sp 78.6% | No healthy subjects. Severity classification not done. LUS not used |
Manfredi et al. [38] | 137 RA patients (59 ILD + , 78 ILD-) | Identifying RA patients with possible ILD using lung sound-based binary classification (presence vs absence of ILD) compared with chest HRCT | FFT | Overall performance: Acc 83.9%, Se 93.2%, Sp 76.9% Performance for velcro-like crackles detection: Acc 67.2%, Se 69.1%, Sp 65.7% | No healthy subjects. Severity classification not done. LUS not used |
Messner et al. [77] | 7 IPF, 16 healthy subjects | Binary classification (healthy subjects vs patients with IPF) | CRNN | F-score 92.4% Se 85.9% | Small number of patients. Manual crackles labelling required. Severity classification not done. LUS not used |
Messner et al. [78] | 5 IPF, 10 healthy subjects | Binary classification (healthy subjects vs patients with IPF) | GRNN | F-score 72.1% Se 71.5% | Small number of patients. Manual crackles labelling required. Severity classification not done. LUS not used |
Ono et al. [79] | 21 IPN, 10 healthy subjects | IPN detectability and severity classification | FFT | Spectral analysis of lung sounds is useful in the diagnosis and evaluation of the severity of IPN | LUS not used |
Pancaldi et al. [17] | 70 RA patients (27 ILD + , 43 ILD-) | Identifying RA patients with possible ILD using lung sound-based binary classification (presence vs absence of ILD) compared with chest HRCT | FFT | Acc 90%, Se 92.6%, Sp 88.4% | No healthy subjects. Severity classification not done. LUS not used |
Santiago-Fuentes et al. [80] | 19 ILD (10 IPF, 9 CPFES) | Binary classification (IPF vs CPFES) | AR model, SNN | Performance evaluation of the neural network: Acc 90.7% to 97.3%, Se 91.8% to 98.3%, Sp 87.5% to 96.3% | No healthy subjects. Severity classification not done. LUS not used |
Sen et al. [81] | 10 obstructive bronchiectasis, 10 ILD, 20 healthy subjects | Binary (healthy vs pathological) and three class (healthy, bronchiectasis and ILD) classifiers | AR model, SVM | Se 85%, Sp 85% Se 100% (three-class classifier) | Small number of patients. Severity classification not done. LUS not used |