A Stacking-based Ensemble Framework for Automatic Depression Detection using Audio Signals

Abstract

Mental illnesses are severe obstacles for the global welfare. Depression is a psychological disorder which causes problems to the individual and also to his/her dependents. Machine learning based methods using audio signals can differentiate patterns between healthy and depressive subjects. These methods can assist health care professionals to detect the depression. Literature in depression detection, based on audio signals, used only single classifier, lacks to take advantage of diverse classifiers. The current work combines predictive capabilities of diverse classifiers using stacking method to detect depression. Audio clips are reordered while a predefined paragraph is being read out, for acoustic analysis of speech. The dataset is created which has extended Geneva Minimalistic Acoustic Parameter Set (eGeMAPS) features, that are extracted using openSMILE toolkit. The normalized feature vectors are given as input to multiple classifiers to give an intermediate prediction. These predictions are combined using a meta classifier to form a final outcome. K-Nearest Neighbours (KNN), Naïve Bayes (NB), Support Vector Machine (SVM), and Decision Trees (DT) classifiers are utilised on the normalized feature vector for intermediate predictions and Logistic Regression (LR) is used as meta classifier to predict final outcome. Our proposed method of using diverse classifiers achieved significant accuracy of 79.1%, precision of 83.3%, recall of 76.9% and F1-score of 80% on our dataset. Results are discussed while using stacking method on our dataset, then compared with various baseline methods also while applying on a publicly available bench marking dataset. Our results showed that combining predictive capability of multiple diverse classifiers helps in depression detection.