Imbalanced Data Classification of Pathological Speech Using PCA, SMOTE, and Expectation Maximization

Imbalance can affect the classification. In this study, a new algorithm to classify the pathology by considering the imbalance between classes has been built. This algorithm used principal component analysis (PCA), synthetic minority oversampling technique (SMOTE), and cluster membership degrees. Both PCA and Expectation-Maximization algorithms are used to give new features combined with the existing features. This proposed method is associated with Support Vector Machine (SVM) and Naive Bayes (NB) for severity classification of UA speech and TORGO speech databases. Another point in this work revealed the importance of articulatory of the TORGO database. The evaluation of this method on the two databases showed the significant results where the increase for TORGO database articulatory features, auditory features, and their combination was respectively 8.03%, 16.68%, and 17.79% compared to SVM performance and 3.04%, 23.67%, and 13.5% compared to the NB performance. The increase for UA speech was 5.13% and 24.63% compared to SVM and NB performance respectively. Additionally, the proposed method has outperformed four well-known imbalanced classification algorithms.