Through the Speech and Vocal Signals Hidden Secrets: An Explainable Methodology for Neurological Diseases Early Detection

Abstract

Neurodegenerative diseases progressively damage brain and nervous systems impairing their functionality. Early diagnosis can improve the efficacy of treatments and patient's life quality. Biomarkers extracted from the human voice can be a simple, efficient, and non-invasive methodology to screen neurodegenerative diseases such as Parkinson's (PD) and multiple sclerosis (MS). Nevertheless, there is still a lack of reliable and clinically approved methodologies required in large-scale patient applications. We define a methodology for features extracted from voice signals as non-invasive indices for early diagnosis of neurodegenerative diseases. We combine and analyze vowels and speech using a set of machine learning (ML) algorithms trained on a combined set of signal features such as acoustic, articulation, and cepstral ones. The methodology has been fully implemented and applied to a dataset of normophonic and pathological voice signals. Experimental results proved that methodology is able to distinguish healthy from pathological voices, with reliable performances, such as accuracy of 97.5%, sensitivity of 98.5%, precision of 97.0%, F1-score of 98.0%, the Matthews correlation coefficient of 0.95, and AUC of 0.98. Finally, the proposed methodology provides explainability tasks for neurological biomarkers identification from speech and vocal features, confirming its reliability. A github repository with data sample and code is available at https://github.com/PatriziaVizza/SpeechAndVocalSignalsAnalysis.

Keywords: Explainability; Machine learning; Neurodegenerative diseases; Neurological biomarkers; Speech; Voice analysis.

Fig. 1

Pipeline of the proposed methodology to report neurological risk indexes. The implemented system processes both speech and vowel signals by extracting relevant features, performs ML analysis and ML explainability, and finally returns a health risk score index. Results explain the correlation between features and (potential) neurological pathology markers

Fig. 2

Workflow of the proposed methodology. The major steps are (i) acquisition and pre-processing of vowel and speech signals, (ii) feature extraction from acoustic, articulatory, and cepstral domains, (iii) application of machine learning algorithms for vocal anomaly classification, (iv) utilization of explainable machine learning techniques to interpret model predictions, and finally (v) longitudinal module to study the progression of the disease

Fig. 3

Receiver operating characteristic (ROC) curves are displayed to visually compare the performance of the five classifiers, i.e., SVM (a), XGBoost (b), DT (c), RF (d), and kNN (e). The area scores are also calculated via stratified three-fold cross-validation to quantify their performance. All models exhibit high area values with SVM and XGBoost demonstrating superior discrimination, while DT, though performing well, presents a slightly lower score, suggesting a potential for overfitting or the need for further hyperparameter optimization

Fig. 4

Confusion matrices for each of the five classification models are reported, showing the comparison between predicted and actual labels. RF and XGBoost demonstrate comparable performance with the most accurate class predictions, where all records of class 1 (i.e., Parkinson) have been associated to the correct class. DT and kNN have the highest false rate. SVM presents a slightly different behavior to RF and XGBoost

Fig. 5

Deep-learning architecture of the proposed model. The architecture is composed of a series of layers which, starting from the input one on top (i.e. “dense”) extract information from input audio files and, in the output layer at the bottom (i.e. “dense_4”, having 3 neurons), predict the probability of their 3 class (i.e., 0 healthy, 1 parkinson, 2 multiple sclerosis)

Fig. 6

Deep model performances in terms of accuracy and loss for both training and test, confusion matrix related to the test set, and overall per-class performance of the model

Fig. 7

Example of force plots for local explainability in healthy (a) and disease (b) instances. In a, a final prediction of $-$1.32, below the base value of 1.732, is shown, resulting in a healthy classification. In b, a prediction of 4.48, significantly above the base value, indicates a disease classification. Both force plots highlight MFCCmedian, MFCCmean, and QVSA as key features, demonstrating their influential positive and negative contributions to the classification

Fig. 8

SHAP summary plot for global explainability. MFCC median and MFCCmean values strongly indicate neurological pathologies, while TVSA and QVSA, though less influential, require local analysis due to their variable SHAP values, with blue (positive) SHAP values signifying pathological predictions and red (negative) values indicating healthy predictions