Error estimation of feature extraction algorithms and weighted classification method for vocal signals

Parkinson disease is a neurodegenerative disorder characterized by a slow but progressive evolution. Even if it mainly involves the motor system, issues with the phonatory system have been noticed. The patient loses full control of the speech apparatus; are noticed uncontrolled repetitions, incorrect articulation of words and a weakening of the voice speech. In recent years, non-invasive techniques based on speech signal processing have been developed for the purpose of early diagnosis and to monitor the effects of pharmacological and neuro stimulation therapies. This thesis work can be considered a primary step of a study focused on improving the models under analysis by increasing the database of the monitored subjects and identifying repeatable patterns using a weighted classification algorithm based on the estimate error of the feature extraction algorithm. To acquire the vocal signals, wearable devices have been used that are able to monitor the subjects not impairing their daily activities. The analyzed signals have been recorded during the repetition of 3 vowels produced by 57 healthy subjects (HE), 67 patients with vocal tract diseases (PA) and 45 parkinsonian patients (PD). From this dataset very unbalanced in terms of both age and gender, a reduced balanced dataset has been extracted, which includes 10 subjects from each class. The aim of the first part of this work has been the processing of the vocal signal to extract parameters that allow to evaluate the stability in frequency(jitter) and amplitude(shimmer) of the sustained vowels and other parameters related to the signal quality, such as harmonic to noise ratio and cepstral peak prominence smoothed (CPPS). To increase the balanced dataset, artificial vowels with known sequences of periods and amplitudes have been generated through a Monte Carlo simulation with the Metropolis-Hastings algorithm. The parameters extracted from the artificial signals have been compared to those extracted from the original signals in order to estimate the error of the feature extraction algorithm. The second part of the work has been focused on the evaluation of reproducibility and repeatability of the obtained measurements. The synthetic vowels previously obtained have been reproduced using a “Head and torso simulator” in an anechoic chamber. The vocal signals produced have been acquired using 3 different measurement chains: - a microphone in air placed in 4 different positions; - a reference phono-meter; - a microphone embedded in an iPhone8; The same parameters have been extracted to evaluate the errors of the feature extraction algorithm by comparing the sequences obtained from these recordings to those extracted from the artificial signals. In the third part of the work, the extracted features have been used to train a weighted logistic regression model to discriminate HE and PA subjects from PD subjects. The combination of features considered by the classifier were the one that had the lowest average relative error. The weights of the features of the algorithm are the reciprocal of the errors obtained in the previous steps. The classification method, considering the original vowels, provided the probability of belonging to HE class with an accuracy of 84.2% and to PA class with an accuracy of 90.0%. This method, using the weights obtained in the previous steps, provided the probability to belong to HE class with an accuracy 93.3% and to PA class with an accuracy of 93.3%