Improving Depression Prediction Accuracy Using Fisher Score-Based Feature Selection and Dynamic Ensemble Selection Approach Based on Acoustic Features of Speech

Abstract

Depression affects over 322 million people, and it is the most common source of disability worldwide. Literature in speech processing revealed that speech could be used for detecting depression. Depressed individuals exhibit varied acoustic characteristics compared to non-depressed. A four-staged machine learning classification system is developed to investigate the acoustic parameters to detect depression. Stage one uses speech recordings from a publicly available and clinically validated dataset DAIC-WOZ. The baseline acoustic feature vector, eGeMAPS, is extracted from the dataset in stage two. Adaptive synthetic (ADASYN) is performed along with data preprocessing to overcome the class imbalance. In stage three, we conducted feature selection (FS) using three techniques; Boruta FS, recursive feature elimination using support vector machine (SVM-RFE), and the fisher score-based FS. Experimentation with various machine learning base classifiers like gaussian naïve bayes (GNB), support vector machine (SVM), k-nearest neighbors (KNN), logistic regression (LR), and random forest classifier (RF) is performed in stage four. The hyperparameters of the classifiers are tuned using the GridSearchCV technique throughout the 10-fold stratified cross-validation (CV). Then we employed multiple dynamic ensemble selection of classifier algorithms (DES) with k=3 and k=5 utilizing the pool of aforementioned four base classifiers to improve the accuracy. We present a comparative study using eGeMAPS features against the base classifiers and the experimented DES classifiers. Our results on the DAIC-WOZ benchmark dataset suggested that K-Nearest Oracles Union (KNORA-U) DES with k=3 has superior accuracy using a subset of 15 features selected by fisher score-based FS than the individual base classifiers.