Vocal feature extraction and classification algorithms for Parkinson's Disease: oversampling method for imbalanced dataset

This thesis work is part of the research field that investigates speech impairments in Parkinson’s Disease (PD) patients, in collaboration with DET and DENERG of Politecnico di Torino. The main goals are the early diagnosis and the evaluation of treatments effectiveness of PD through the processing and classification of voice signals. The available voice dataset was acquired in previous years at Centro Parkinson of Molinette hospital (TO) using a microphone in air and a throat contact microphone. The monitored patients belong to 3 different classes: healthy (HE), PD and pathological (PA) subjects and the tasks they performed are sustained vowel /a/, reading and free speech. The first part of the thesis is focused on the signals pre-processing to subdivide voice signals into silence, voiced and unvoiced frames, according to three basic features: rms (root-mean-square value), fundamental frequency F0 and HNR (Harmonic-to-Noise Ratio). Subsequently, the descriptive statistics of F0, HNR, rms and CPPS (Cepstral Peak Prominence Smoothed) have been extracted only for the voiced frames. Moreover, period (Jitter) and amplitude (Shimmer) stability parameters have been extracted from the signals related to the sustained vowel /a/. The available dataset is characterized by a strong imbalance in age and gender of the subjects, therefore it has been necessary to assess the effect of these factors on the extracted parameters through a non-parametric statistical test. Afterwards, to balance the dataset, an attempt has been made to over-sample the minority class. In this thesis the method used is SMOTE (Synthetic Minority Over-sampling TEchnique) that has been applied on healthy elderly subjects. First, the reliability of this technique has been assessed by comparing the distributions of the original and the synthetic features, produced by SMOTE. Later, SMOTE has been applied to the 10 oldest healthy subjects with a 300% oversampling rate, thus attaining the features of 40 subjects for HE class. At the same time, 40 patients have been selected from PA and PD classes, excluding the oldest, in order to obtain comparable ages between the 3 classes. The classification of subjects in the 3 classes, on the basis of the extracted parameters, has been performed both on the original dataset and on the balanced dataset, after oversampling, for each task and for both microphones. The implemented classifiers are the decision tree algorithm and the logistic regression for binary classifications (HE vs PD and PA vs PD). Furthermore, decision-tree algorithm has been trained for the multi-class classification. To avoid overfitting and thus obtain more realistic performance, k-fold cross validation has been used, with k=5. The implemented metrics, to evaluate the classification models performance, are accuracy, sensitivity, specificity and AUC (Area Under the Curve). With decision trees, accuracies up to 95% for the microphone in air and 90% for the contact microphone have been achieved, while the logistic regression has provided accuracies up to 95% for the microphone in air and 96% for the contact microphone. The performance of both classifiers is higher on the balanced dataset, after SMOTE, than on the original dataset. In the future, the dataset could be extended by means of a smartphone app, which allows speech material to be remotely collected, and through traditional acquisition methods, when it will be possible to return to the hospital.