Towards the differential evaluation of parkinson's disease by means of voice and speech processing

• Autores: Laureano Moro Velázquez
• Directores de la Tesis: Juan Ignacio Godino Llorente (dir. tes.)
• Lectura: En la Universidad Politécnica de Madrid ( España ) en 2018
• Idioma: español
• Tribunal Calificador de la Tesis: Javier Macías Guarasa (presid.), José Luis Blanco Murillo (secret.), Francisco Grandas Pérez (voc.), Mekyska Jiri (voc.), Rubén San Segundo Hernández (voc.)
• Programa de doctorado: Programa de Doctorado en Ingeniería de Sistemas y Servicios para la Sociedad de la Información por la Universidad Politécnica de Madrid
• Menciones:
  • Internacional
• Calificación: Sobresaliente Cum Laude
• Materias:
  • Matemáticas
    • Ciencia de los ordenadores
      • Inteligencia artificial
    • Estadística
      • Análisis de datos
  • Ciencias médicas
    • Medicina interna
      • Neurología
  • Lingüística
    • Lingüística sincrónica
      • Fonética
• Enlaces
  • Tesis en acceso abierto en: Archivo Digital UPM
• Resumen

  • Parkinson’s Disease (PD) is a neurodegenerative condition that affects to 1% of population over the age of 60 in industrialized countries. This disease seriously affects a patient’s independence and motor capabilities, having a considerable impact on their daily activities as it advances.

    Early detection can potentially slow the progression of PD but, unfortunately, the required period of clinical diagnosis ranges from months to years. Therefore, new objective and reliable tools are needed to support the differential diagnosis of the disease and to reduce this time.

    The analysis of a patient’s speech has demonstrated to provide relevant information about the presence of the disease and, consequently, is a possible source of features to be used in diagnosis systems. Speech, an ability that is almost universal, involves coordination and precision of movements in mainly the laryngeal and articulatory muscles. The purpose of this thesis is to propose and study different approaches to support clinical diagnosis of PD employing speech as the object of analysis.

    In this thesis, five sets of experiments are carried out, each one containing new approaches aimed to detect the presence of the disease in the speech of idiopathic PD patients and controls from three different databases. Four of these experiments focus on the analysis of articulatory aspects while the fifth employs phonatory features and a final combination of phonatory and articulatory information into a single approach. In these approaches, several state-of-the-art speaker and speech recognition technologies are employed in a different scenario: the automatic detection of PD from speech.

    Several known feature families such as MFCC, PLP or LPC and new features based on the use of the modulation spectra are analyzed. Moreover, different speech frame selection techniques are proposed, such as allophonic distillation and acoustic landmark distillation, providing certain specific speech segments that are of interest to the purposes of this work. The main classification techniques employed are GMM-UBM and i-Vectors-GPLDA, along with new schemes such as the forced GMM.

    As a consequence of the analysis of the proposed approaches, the influence of PD in these specific segments is examined, allowing to extract conclusions about the functioning of the parkinsonian dyshartric speech. These segments are phonetic groups related to the narrowing of the vocal tract or the use of the glottal source, such as fricatives, plosives or vowels, or include mainly transitions between phonetic units, such as the beginning of a burst or the end of a vowel, more related with the coordination of the articulators.

    The best accuracy results in the detection of PD achieved with the proposed methodologies reach values ranging from 85% to 94% with Area Under the Curve between 0.91 and 0.99 depending on the database. These results are obtained largely by employing approaches based on the frame selection proposed techniques: allophonic and acoustic landmark distillations. In the same manner, it is concluded that the discriminatory properties of the proposed phonatory approaches to automatically detect PD are quite limited in comparison with the analyzed articulatory approaches.

    Results suggest that PD affects the movements related to all of the studied articulatory segmental groups but has a clearer influence in the consonants with a greater narrowing of the vocal tract, mainly plosives.

    Finally, the new proposed methodologies demonstrate their ability to support the differential diagnosis of PD during a patient’s clinical assessment and are a step forward in the speech-based diagnosis systems for PD employing articulatory aspects.