Leveraging Deep Learning based Feature Extraction Techniques for Improved Audio-Based Depression Detection

A significant number of people across the globe are affected by Depressive disorder due to various reasons. The health care practitioners use several traditional methods in their practice to diagnose depression which includes clinical assessment and scoring through questionnaires however those methods proved to have pitfalls in inaccurate diagnosis. In this paper, audio data is utilised for the detection of depression. This paper is related to the Sustainability Development Goal-SDG 3 by diagnosing depressive disorders and taking preventive actions to create the healthy mental wellbeing and social atmosphere. The approach employed in this research is centered on leveraging feature extraction techniques based upon deep learning for identifying depressed subjects. Investigations are conducted within patients and controls data obtained from one of the popular publicly available datasets named EDAIC-WOZ dataset. MFCC (Mel Frequency Cepstral coefficients) features are extracted and corresponding graphs are plotted followed with using Local binary pattern and training with various classifiers (Support Vector Machine, Multi-Layer Perceptron, Decision tree) and evaluate accuracy. As per the results, the Decision Tree Classifier confirms the paramount accuracy at 0.939, surpassing both SVM and MLP classifiers accuracy values.