Voice Disorder Detection via an m-Health System: Design and Results of a Clinical Study to Evaluate Vox4Health

Abstract

Objectives. The current study presents a clinical evaluation of Vox4Health, an m-health system able to estimate the possible presence of a voice disorder by calculating and analyzing the main acoustic measures required for the acoustic analysis, namely, the Fundamental Frequency, jitter, shimmer, and Harmonic to Noise Ratio. The acoustic analysis is an objective, effective, and noninvasive tool used in clinical practice to perform a quantitative evaluation of voice quality. Materials and Methods. A clinical study was carried out in collaboration with medical staff of the University of Naples Federico II. 208 volunteers were recruited (mean age, 44.2 ± 13.9 years), 58 healthy subjects (mean age, 36.7 ± 13.3 years) and 150 pathological ones (mean age, 47 ± 13.1 years). The evaluation of Vox4Health was made in terms of classification performance, i.e., sensitivity, specificity, and accuracy, by using a rule-based algorithm that considers the most characteristic acoustic parameters to classify if the voice is healthy or pathological. The performance has been compared with that achieved by using Praat, one of the most commonly used tools in clinical practice. Results. Using a rule-based algorithm, the best accuracy in the detection of voice disorders, 72.6%, was obtained by using the jitter or shimmer value. Moreover, the best sensitivity is about 96% and it was always obtained by using jitter. Finally, the best specificity was achieved by using the Fundamental Frequency and it is equal to 56.9%. Additionally, in order to improve the classification accuracy of the next version of the Vox4Health app, an evaluation by using machine learning techniques was conducted. We performed some preliminary tests adopting different machine learning techniques able to classify the voice as healthy or pathological. The best accuracy (77.4%) was obtained by the Logistic Model Tree algorithm, while the best sensitivity (99.3%) was achieved using the Support Vector Machine. Finally, Instance-based Learning performed the best specificity (36.2%). Conclusions. Considering the achieved accuracy, Vox4Health has been considered by the medical experts as a “good screening tool” for the detection of voice disorders in its current version. However, this accuracy is improved when machine learning classifiers are considered rather than the rule-based algorithm.