Table 1.
Reference | Year | Topic | Data and Cohort | Recording Device | ML Models Used | Data Processing Methods | KPIs |
---|---|---|---|---|---|---|---|
[8] | 2020 | Cough Detection | Private-110 subjects (various resp. diseases) | Various Microphones | XGBoost, RF, DT | Feature Extraction from cough events | Acc: 93% Sens: 97% Spec: 95 |
[9] | 2020 | Disease Identification: COVID-19 | Private-2660 subjects | Web App, Various Microphones | CNN w/1 Poisson biomarker and 3 pre-trained ResNet50 | MFCCs | AUC: 97% |
[10] | 2020 | Disease Identification: COVID-19 | ESC-50, Audioset-Cohort n.s. | Various Microphones | CNN, VGG16 | MFCCs | Accuracies: COVID/Non-COVID: 70% Cough/Non-Cough: 90% |
[11] | 2021 | Disease Identification: COVID-19 | Private-8380 positive, 6041 negative instances | N/A | 2D CNN | EMD and MFCCs | AUC: 97% |
[12] | 2021 | Disease Identification: COVID-19 | Coswara, Virufy-Cohort n.s. | Web App, Various Microphones | SVM, KVM, RF | Common Short Term Features and MFCCs | AUC: 90% |
[13] | 2021 | Disease Identification: COVID-19 | Private-240 acoustic data—60 normal, 20 COVID-19 subjects | Smartphone | LSTM (RNN) | Spec. Centroid, Spec. roll-off, ZCS, MFCC (+ΔΔ) | Acc: 97%, F1: 97.9% |
[6] | 2021 | Disease Identification: COVID-19 | Private-Cohort n.s. | Crowdsourced audio recordings, Various Microphones | XGBoost | Feature Extraction | Acc: 97% |
[14] | 2020 | Cough Detection, Disease Identification: COVID-19 | ESC-50 and Private-543 Recordings (96 bronchitis, 130 pertussis, 70 COVID-19, 247 Normal) | N/A | 1 CNN for Cough Classification and 3 CNNs for COVID-19 detection | Mel-spectrograms to images | Accuracies: Cough Detection: 95.5% COVID-19 Identification: 92.64% |
[15] | 2021 | Disease Identification: Bronchitis, Asthma, COVID-19, Healthy | N/A | N/A | Fully connected NN layer | Questionnaire and Cough Embeddings | Acc: 95.04% |
[16] | 2022 | Cough Detection, Disease Identification: COVID-19 | Virufy-Cohort n.s. | N/A | DNN | Windowing and Feature Extraction | Acc: 97.5% |
[17] | 2022 | Cough Detection | Corp Dataset-42 volunteers (18 CAP, 4 asthma, 17 COPD, 3 other resp. illness) | Digital Recorder | CNN | MFCCs | Acc: 99.64%, IoU: 0.89 |
[18] | 2022 | Cough Detection | Public/ Private-3228 cough and 480,780 non-cough sounds | Various Microphones | GB Classifier | Feature Extraction and manual selection | Acc: 99.6% (Validation only on hospital data from children) |
[19] | 2022 | Cough Detection, Disease Identification: COVID-19 | Cambridge, Coswara, Virufy, NoCoCoDa - Cohort n.s. | Various Microphones | Adaboost, MLP, XGBoost, Gboost, LR, K-NN, HGBoost to MCDM | Feature Extraction | Acc: 85% |
[5] | 2022 | Disease Identification: COPD, AECOPD | Private-177 volunteers (78 COPD, 86 AECOPD, 13 not used) | Various Microphones | N/A | Feature Extraction | ROC Curve: 0.89, agreement w/ clinical study |
[20] | 2013 | Symptom Identification: Wet Cough, Dry Cough | Private-78 Patients | High-end Microphones | LR | Feature Extraction | Sens: 84%, Spec: 76% |
[21] | 2020 | Cough Detection | Private-26 healthy participants | Smartphone | K-NN, DT, RF, SVM | Feature Extraction and Selection | F1 Score: 88% |
[22] | 2019 | Cough Detection | Private-20 min of cough sounds | N/A | Hidden Markov Models | Single and Multiple Energy Band | AUC: 0.844 |
[23] | 2020 | Cough Detection | Private-94 adults | Smartphone | CNN | Mel spectrograms | Accuracies: Cough Detection: 99.7% Sex Classification: 74.8% |
[24] | 2018 | Cough Detection, Disease Identification: Croup | Private-56 croup and 424 non-croup subjects | Smartphone | SVM and LR | Feature Extraction | Sens 92.31% Spec: 85.29% Croup classification Acc: 86.09% |
[25] | 2020 | Cough Detection | N/A | N/A | SVM and K-NN | Feature Extraction | Accuracies: K-NN: 94.51%, SVM: 81.22% |
[26] | 2020 | Symptom Identification: Wet Cough, Dry Cough | Private-5971 coughs (5242 dry and 729 wet) | Smartphone | RF | Feature Extraction (Custom and OpenSmile) | Acc: 88% |
[27] | 2021 | Disease Identification: Heart-failure, COVID-19, Healthy | Private-732 patients (241 COVID-19, 244 Heart-failure, 247 Normal. | Smartphone | K-NN | DNA Pattern Feature Generator, mRMR Feature Selector | Acc: 99.5% |
[28] | 2020 | Cough Detection, Disease Identification: Pertussis, Bronchitis, Bronchiolitis | ESC-50, Audioset-Cohort n.s. | Various Microphones | CNN | Mel-spectrograms | Accuracies: Disease Identification: 89.60%, Cough Detection: 88.05% |
[29] | 2019 | Cough Detection, Symptom Identification: Productive Cough, Non-productive Cough | Private-810 events: 229 non-productive cough, 74 productive cough, and 507 other sounds | Various Microphones | DLN | FFT and PCA | Acc: 98.45% |
[30] | 2017 | Disease Identification: Croup | Private-364 patients (43 Croup, 321 non-croup | Smartphone | LR and SVM | MFCCs and CIFs | Acc: 98.33% |
[4] | 2020 | Disease Identification: Asthma | Private-997 asthmatic, 1032 healthy sounds | Smartphone | GMM - UBM | MFCCs and CQSSs | Acc: 95.3% |
[31] | 2021 | Disease Identification: COVID-19 | Coswara-Cohort n.s. | Smartphone | ResNet50 | Feature Extraction | Acc: 97.6% |
[32] | 2022 | Disease Identification: COVID-19 | Coswara, Virufy, Cambridge and private-Cohort n.s. | Various Microphones | Most popular supervised models | Feature selection and extraction | Best Acc: Random Forest: 83.67% |
[33] | 2022 | Disease Identification: Tuberculosis | Private-16 TB and 35 non-TB patients | Various Microphones | LR, SVM, K-NN, MLP, CNN | Feature Extraction | Best Acc: LR: 84.54% |
[34] | 2021 | Disease Identification: COVID-19 | Virufy, NoCoCoDa-Cohort n.s. | Various Microphones | SVM, LDA, K-NN | Feature Extraction and SFS feature selection | Best Acc: K-NN: 98.33% |
[7] | 2021 | Cough Detection | ESC-50-50 cough recordings | Various Microphones | CNN | Mel spectrograms and Data Augmentation | Acc: 98% |
[35] | 2021 | Disease Identification: COVID-19 | Coswara, COVID-19 Sounds-Cohort n.s. | Various Microphones | Contrastive Learning | Contrastive Pre-training: Feature Encoder w/ Random Masking | Acc: 83.74% |
[36] | 2019 | Disease Identification: Asthma, Healthy | Private-89 children for each cohort, 1992 healthy and 1140 asthmatic cough sounds | Smartphone | Gaussian mixture model | Downsampling and multidimensional Feature Extraction (MFCCs and CQCCs) | Sens: 82.81%, Spec: 84.76%, AROC: 0.91 |
Note. ML models: SVM = Support Vector Machine; K-NN = K-Nearest Neighbors; DT = Decision Trees; RF = Random Forest; NN = Neural Network; CNN = Convolutional Neural Network; RNN = Recurrent Neural Network; DLN = Deep Learning Networks; MLP = Multilayer Perceptron; LSTM = Long Short-Term Memory; GB classifier = Gradient Boosting classifier; HGBoost = Hyperoptimized Gradient Boosting; LR = Linear Regression; GMM = Gaussian Mixture Model; UBM = Universal Background Models; LDA = Linear Discriminant Analysis. Data Processing Methods: MFFCs = Mel-Frequency Cepstral Coefficient; mRMR = minimum Redundancy − Maximum Relevance; FFT = Fast Fourier Transform; PCA = Principal Component Analysis; CIF = Cochleagram Image Features; CQCC = Constant-Q Cepstral Coefficients. Metrics: Acc = Accuracy; Sens = Sensitivity; Spec = Specificity; AUC = Area Under Curve; IoU = Intersection over Union; ROC Curve = Receiver Operating Characteristic Curve; AROC = Area under the ROC curve.