ALGORITHM DEVELOPMENT FOR PARKINSON'S DISEASE DETECTION BASED ON SPEECH ANALYSIS Artificial Intelligence applied to disease diagnosis and patient follow-up

Automatic assessment of speech impairment is a cutting edge topic in Parkinson's disease (PD) detection. Patients usually face loss of prosody, volume and clarity, which results in a dysfunction of the different levels involved in speech production. This condition is clinically referred as dysarthria and is characterized by alterations in speed, volume, tone, range or precision of movements necessary for voice control. Despite a corroborated methodology is currently employed by the clinician to asses the presence and the level of Parkinson's disease, the development of an automatic tool able to reduce the operator-dependency is still an hot research topic. Furthermore, as far as we know, automatic methodologies for PD detection make mainly use of vocalization exercises, and the set of features to be extracted is demonstrated being task-dependent. Therefore, parameters to be used for utterances repetition, as in the current study, may differ from the latest ones. Consequently, the possibility of extracting useful information from word databases is still being investigated, and a set of features whose effectiveness is validated for the specific task has not been created yet. In addition, most of the research studies focus on English native speakers, leaving as an open question the possibility to extend the features importance to other language analysis. This thesis has deepened the idea that features extracted from vocal samples of patients with PD contain information about the pathological condition of the individual, and therefore can be used to facilitate the diagnosis and the follow-up stages. Native Spanish speakers have been used to validate the features importance in neo-Latin languages. The developed software consists into a feed-forward artificial neural network and a long short-term memory artificial recurrent neural network, fed with features extracted from the pre-processed signals. The parameters employed include either classical features, whose effectiveness in PD vocal analysis is corroborated, or novel characteristics, whose relevancy has been investigated in this dissertation. Furthermore, a post-hoc analysis of the trained model has been carried out in order to increase the interpretability of the LSTM network, and to extract further information about the importance of each parameter in voice impairment recognition of PD patients.