Publicação
Identification of voice pathologies in an elderly population
| Resumo: | Ageing is associated with an increased risk of developing diseases, including a greater pre- disposition to develop diseases such as Sepsis. Also, with ageing, human voices undergo a natural degradation gauged by alterations in hoarseness, breathiness, articulatory ability, and speaking rate. Nowadays, perceptual evaluation is widely used to assess speech and voice impairments despite its high subjectivity. This dissertation proposes a new method for detecting and identifying voice patholo- gies by exploring acoustic parameters of continuous speech signals in the elderly popula- tion. Additionally, a study of the influence of gender and age on voice pathology detection systems’ performance is conducted. The study included 44 subjects older than 60 years old, with the pathologies Dyspho- nia, Functional Dysphonia, and Spasmodic Dysphonia. In the dataset originated with these settings, two gender-dependent subsets were created, one with only female samples and the other with only male samples. The system developed used three feature selection methods and five Machine Learning algorithms to classify the voice signal according to the presence of pathology. The binary classification, which consisted of voice pathology detection, reached an accuracy of 85,1%±5,1% for the dataset without gender division, 83,7%±7,0% for the male dataset, and 87,4%±4,2% for the female dataset. As for the multiclass classifica- tion, which consisted of the classification of different pathologies, reached an accuracy of 69,0%±5,1% for the dataset without gender division, 63,7%± 5,4% for the male dataset, and 80,6%±8,1% for the female dataset. The obtained results revealed that features that describe fluency are important and discriminating in these types of systems. Also, Random Forest has shown to be the most effective Machine Learning algorithm for both binary and multiclass classification. The proposed model proves to be promising in detecting pathological voices and identifying the underlying pathology in an elderly population, with an increase in its performance when a gender division is performed. |
|---|---|
| Autores principais: | Silva, Nisa Mafalda Lagos |
| Assunto: | Voice Pathology Detection Acoustic Parameters Machine Learning Elderly Population |
| Ano: | 2022 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade Nova de Lisboa |
| Idioma: | inglês |
| Origem: | Repositório Institucional da UNL |
| Resumo: | Ageing is associated with an increased risk of developing diseases, including a greater pre- disposition to develop diseases such as Sepsis. Also, with ageing, human voices undergo a natural degradation gauged by alterations in hoarseness, breathiness, articulatory ability, and speaking rate. Nowadays, perceptual evaluation is widely used to assess speech and voice impairments despite its high subjectivity. This dissertation proposes a new method for detecting and identifying voice patholo- gies by exploring acoustic parameters of continuous speech signals in the elderly popula- tion. Additionally, a study of the influence of gender and age on voice pathology detection systems’ performance is conducted. The study included 44 subjects older than 60 years old, with the pathologies Dyspho- nia, Functional Dysphonia, and Spasmodic Dysphonia. In the dataset originated with these settings, two gender-dependent subsets were created, one with only female samples and the other with only male samples. The system developed used three feature selection methods and five Machine Learning algorithms to classify the voice signal according to the presence of pathology. The binary classification, which consisted of voice pathology detection, reached an accuracy of 85,1%±5,1% for the dataset without gender division, 83,7%±7,0% for the male dataset, and 87,4%±4,2% for the female dataset. As for the multiclass classifica- tion, which consisted of the classification of different pathologies, reached an accuracy of 69,0%±5,1% for the dataset without gender division, 63,7%± 5,4% for the male dataset, and 80,6%±8,1% for the female dataset. The obtained results revealed that features that describe fluency are important and discriminating in these types of systems. Also, Random Forest has shown to be the most effective Machine Learning algorithm for both binary and multiclass classification. The proposed model proves to be promising in detecting pathological voices and identifying the underlying pathology in an elderly population, with an increase in its performance when a gender division is performed. |
|---|