Design and Implementation of method to detect depression using Sentiment Analysis

Abstract

This study presents a novel multimodal technique for the exposure of depression newlineand the assessment of its severity levels by integrating audio and text data. The newlinemethodology combines a 2D Convolutional Neural Network (2DCNN) for audio newlineinvestigation and a Bidirectional Gated Recurrent Unit (BiGRU) for text processing. This newlinemethod allows for a comprehensive exploration of depression-related information. The newlinekey steps of the methodology include Audio Processing where audio data is processed newlineusing a 2D CNN layer, extracting spatial features from audio features. The Rectified newlineLinear Unit (ReLU) activation function familiarizes non-linearity. Text Processing newlineincludes textual data undergoes processing with a Bidirectional Gated Recurrent Unit newline(BiGRU), which excels in sequential data analysis, capturing contextual information from newlinetext in both forward and backward directions. Features extracted from both audio and text newlinemodalities are combined into a unified feature vector. This combined representation newlineincludes flattened outputs from the audio CNN, the text BiGRU, and a scalar PHQ-8 score newlineinput. The combined features are passed through two separate Dense layers. The first newlineDense layer produces a binary classification output, indicating the existence or newlinenonexistence of depression, using sigmoid activation. The second Dense layer provides a newlinelevels classification output, predicting the severity of depression using softmax activation newlineon a scale. The combination of text and audio data at the feature level enhances newlineclassification performance, addressing both binary and levels classification of depression. newlineExperimental outcomes using the DAIC-WOZ database determines the efficiency of the newlinemethod. Various audio descriptors, including MFCC, chroma features, Spectral density, newlineand Mel spectrograms, were evaluated, with combined MFCC, chroma, Spectral density, newlineand Mel spectrograms yielding superior performance. For transcript features, pre-trained newlinemorphological representations such as Bert outperformed ELMo in en