Table 7 A comparison of acoustic data/features papers
Article | Main idea | Tools | Applied algorithms | Advantages | Disadvantages |
|---|---|---|---|---|---|
DCNN-based model for pathological speech detection | • MATLAB | • DCNN • SpecAugment • WOA | • High precision • High sensitivity • High accuracy • High specificity • High F1-score | • Not evaluating a large hybrid dataset • Lack of multilingual data • Not investigating other classifiers or forms of DCNN | |
ML-based approach using speech signal analysis | • Not mentioned | • SVM • CNN • GMM • MLP • Adam optimizer • Expectation Maximization | • High F1-score • High accuracy • High sensitivity • High specificity • High precision | • Low scalability | |
DL and ML-based PD diagnosis via speech signals | • Not mentioned | • CNN • RF | • High accuracy • High precision • High recall • High AUC • High F1-score • High specificity | • Low scalability | |
Ensemble-based model for PD detection using voice signals | • Not mentioned | • DNN • EOFSC | • High accuracy | • Not analyzing parameters such as F1-score and precision | |
PD detection using DL and ACSA algorithm | • Python (TensorFlow) | • ACSA • AutoEncoder | • High accuracy • High sensitivity • High specificity | • Not investigating unsupervised classification methods • Lack of identification of scrunched feature vectors | |
Acoustic-based PD detection using IGKNN algorithm | • Not mentioned | • KNN | • High accuracy • High recall | • High cost • Poor handling of multi-dimensional features | |
Optimized ResNet50 model for PD diagnosis | • Not mentioned | • ResNet50 • GDABC | • High accuracy | • High time consumption | |
Two-phase ML-based system for early PD diagnosis | • Not mentioned | • LightGBM | • High accuracy • High AUC • High precision • High F1-score | • Not optimizing the hyperparameters | |
TL-based approaches for PD detection via speech | • MATLAB | • KNN • SVM • FT & KNN • FT&SVM • TSTL | • High sensitivity • High accuracy • High specificity | • Lack of multimodal data | |
Speech-based PD diagnosis using a neural network | • Not mentioned | • PNN | • High accuracy | • Lack of mobile apps for PD detection | |
PD detection using DNN on voice data | • Not mentioned | • DNN | • High accuracy | • Low scalability | |
Feature correlation–driven ensemble model for early PD diagnosis | • Python | • MLP • NB • KNN • SVM • DT | • High accuracy • High AUC | • lack of clinical validation | |
ML-based feature selection for PD diagnosis | • Not mentioned | • SVM | • High accuracy | • Low scalability • Low quality of the recording device | |
CNN on speech recordings for PD diagnosis | • Not mentioned | • ResNet50 • Xception • SGD | • High accuracy • High specificity • High sensitivity • Language independence • High ROC (AUC) | • Low scalability • Not considering the drug effects | |
DL-based PD diagnosis via feature extraction | • MATLAB | • DSL • L1R&FS • SVM • EGSAE • IMC • WFM-DSS | • High accuracy | • Not exploring methods to rebuild deep samples • Not applying more DNN architectures | |
End-to-end DL model for PD diagnosis | • Python (Scikit-learn, librosa) • NeuroSpeech | • 2D-CNN • 1D-CNN | • High accuracy • High specificity | • Low scalability | |
Optimized DCNN for PD & cleft lip and palate | • MATLAB | • DCNN • ChOA • IPChOA • SpecAugment | • High specificity • High accuracy • High F1-score • High sensitivity • High precision | • Lack of other languages in the dataset • Not investigating other CNNs or GAN • Not evaluating other non-pathology-related cues | |
Speech-based PD diagnosis using SVM | • Not mentioned | • SVM • FFT • DCT • AutoEncoder | • High accuracy • High recall • High F1-score | • Low scalability • Lack of phonation, prosody, and speech rhythm evaluation | |
Stacking ensemble model for PD diagnosis | • Python | • XGBoost • RF • LR • Stacking ensemble technique | • High accuracy • High F1-score • High precision • High recall | • Low scalability • Lack of comprehensive clinical data | |
PD diagnosis using ANN and speech features | • Statistical software | • ANN • Levenberg-Marquardt • Back-propagation algorithm | • High accuracy • High sensitivity • High AUC • High specificity | • Not evaluating additional languages • Low scalability | |
PD diagnosis via ML and speech signals | • Not mentioned | • mRMR • RF • LR • KNN • DNN | • High F1-score • High accuracy • High MCC | • Unclear data accuracy and privacy • Lack of patient usability and convenience evaluation | |
Evaluation of ML classifiers for the detection of PD | • WEKA | • J48 • NB-tree • MPNN • SVM | • High accuracy | • Low scalability • Not evaluating other advanced ML algorithms • Not evaluating feature selection techniques | |
Ensemble voting for PD detection via acoustic data | • Python | • LR • Ridge classifier • SGD • PAC • KNN • Extra tree • DT • SVC • Gaussian NB • AdaBoost • Bagging classifier • RF • Gaussian process classifier • GB • LDA • QDA • XGBoost • MLP | • High accuracy • High precision • High recall • High AUC • High F1-score • High specificity | • Low scalability | |
ML approach for early PD detection via healthcare decision-making | • R programming | • NB • ANN • DT • RF | • High accuracy | • Lack of applying DL methods • Motor symptoms were not considered | |
PD classification based on audio data using ELM | • MATLAB | • Short-time Fourier transform • ELM • CNN | • Low training time • Low cost • High accuracy | • Low scalability | |
ML-based voice analysis for PD detection | • Not mentioned | • XGBoost • CNN | • High accuracy • High specificity • High sensitivity • High AUC (ROC) | • Only acoustic data intended • Only one of the datasets determines gender | |
Assessing ML reliability for PD voice-based diagnosis | • MATLAB | • SVM • Relief-F algorithm | • High recall • High accuracy | • Low scalability | |
DNN-based classification of PD and SCD using speech data | • Not mentioned | • DNN • Patchout faSt spectrogram transformer | • High AUC (ROC) • High accuracy • High specificity • High sensitivity • Non-invasive screening method | • Low scalability | |
PD diagnosis using speech features | • Python (Scikit-learn) | • SVM • XGBoost • RF • KNN • DT • LR • PCA • SMOTE-ENN | • High accuracy • High F1-score • High AUC (ROC) | • Low scalability | |
ML-based acoustic analysis for PD detection | • R programming • Python | • CNN • RF • LR | • High AUC (ROC) | • Not checking the health status of HC • Low scalability | |
Audio-based PD detection via ML and DL | • Python • Praat | • SVM • DNN • Adam optimizer • HGSA | • High specificity • High sensitivity • High accuracy • High MCC | • Not examining the severity of PD in patients • Low scalability |
