Differential diagnosis support tools for adductor spasmodic dysphonia and muscular tension dysphonia

GASTÓN SCHLOTTHAUER; MARÍA EUGENIA TORRES; HUGO LEONARDO RUFINER; MARÍA CRISTINA JACKSON MENALDI

San Pablo, Brasil

Congreso; XIX ENT World Congress IFOS 2009; 2009

INTERNATIONAL FEDERATION OF ORL SOCIETIES

Adductor spasmodic dysphonia (AdSD) and muscle tension dysphonia (MTD) are two voice disorders with similar characteristics. Many reasons support the idea that AdSD is a neurological disease, requiring botulinum toxin A injections or surgical treatments. MTD is a functional disorder correctable with voice therapy. Usually they can be properly differentiated only by experienced voice clinicians.The importance of a correct diagnosis of these two disorders is critical at the treatment selection moment. In the present work we propose automatic non invasive tools for clinical diagnosis support.We use normal and pathological sustained vowels /a/ and synthetic voices. For classification, we use eight acoustic parameters in two pattern recognition methods: neural networks (NN) and support vector machines (SVM). Looking for a dimensional reduction for visualization purposes, we propose a tool based on empirical mode decomposition (EMD) using six parameters related with spectral properties of the EMD voice decomposition.In the present problem, NN perform better than SVM. Different NN architectures were considered. Using a NN with 14 hidden units, AdSD obtained a 95.24% of correct classifications (CC). MTD reached an 86.67% of CC with 16 hidden units, while normal voices were 100% recognized. In case of pathological vs. normal voices we obtained a 98.94% of CC, overcoming the best reported result (96.5%). Applying EMD and a k-nearest neighbor classifier with normal and pathologic synthetic voices, we have obtained a 99% of CC; while in the case of real voices the percentage of CC was 93.40%. Using three dimensions it is possible to distinguish by visual inspection normal, AdSD, and MTD voices.The results indicate that EMD and pattern recognition based tools are useful for the discrimination between normal and pathological voices, suggesting that it could be possible to develop an automatic tool to help in the diagnosis. Future works will include the study of a wider data base of real signals, in collaboration with voice pathologists.