Classification-based screening of Parkinson’s disease patients through voice signal

In this paper a classification algorithm for Parkinson's Disease screening is proposed. Code executes the processing of specific voice signals recorded by healthy and ill subjects. In the direction of a future implementation and validation in a home telemonitoring system, the algorithm has been built with the objective to serve as a screening tool for the precocious directing of subjects with high risk of neurological diseases to instrumental exams. In fact, in several neurological disorders, such as Parkinson's disease, motor impairments of vocal apparatus arise earlier than postural and ambulatory symptoms. In a home telemonitoring system, in which hardware would consist in a voice recorder (that could be a simple smartphone) and a server for the web platform, data would be acquired and instantly stored on a platform for their processing through machine learning algorithms and to be viewed by specialists. For this purpose, a fully automatic process is needed. Therefore, in this work, audio-preprocessing and features computation are completely performed automatically, using Matlab. Final models have been trained in Matlab environments from Weka's libraries. The family of developed models are trained with different type of phonations, from simple vowels to complex sounds, for a wider and more efficient analysis of vocal apparatus motor impairments. Moreover, dataset was 612 observation large, that is significantly above the mean size of similar works using simple phonations only. For a deeper analysis, different groups of parameters have been tested and cepstral features have been found to be optimal for classification and made up the big part of final algorithm. Developed models are part of the K-Nearest Neighbor family, thus, available for implementation in web platform. Finally, obtained models have shown high accuracies on the whole dataset, reaching values comparable with the literature but with more stability (standard deviation less than 1%). These results have been confirmed in the last validation session in which models have been exported and validated with 25% of data, reaching a best performance with a true positive rate of 98% and a true negative rate of 87%.