A Multi-model Framework for Evaluating Type of Speech Samples having Complementary Information about Parkinson's Disease

Recent research points out that Parkinson's disease (PD) patients show different symptoms including rigid muscles, slowed movement and tremor. However, dysphonia-changes in speech and articulation-is considered the most significant precursor as 90% of PD patients posses speech impairments. Multiple types of speech data have been used for PD detection. The multiple types of speech data includes different types of voice or speech samples for each subject, e.g., speech recordings of numbers, words, some short sentences and vowel phonations. Some of the previous studies pointed out that number samples show better performance while other concluded that vowel samples have complementary information about PD. Hence, this is still an open question that which type of speech samples contain complementary information about PD. Moreover, the drawback in these studies is that their methodology is solely based on one type of machine learning model. In this paper, we propose a more robust methodology based on multi-model framework which uses multiple types of machine learning models having diverse nature. For final evaluation of the multi-model framework, we propose to use two different evaluation criteria i.e., mean and majority voting. Based on the multi-model methodology, our study concludes that vowel samples posses complementary information for PD. Additionally, to further validate the findings of our study, we further used two other evaluation metrics namely Receiver Operating Characteristic (ROC) chart and area under the curve (AUC). The results of ROC charts and AUC further validated the fact that vowel samples possess complementary information regarding PD.