Deep Learning Multimodal Ensemble Techniques for Detecting the Depression Level using Multitudinous Selective Features

In this modern era, massive population of this world is falling prey to mental illness and psychological diseases rapidly along with other physical diseases. Among all, depressive disorder is becoming major challenging and serious psychological problem. Due to a lack of initial treatment options and resources for depression detection, millions of people struggle with mental illness. It is the primary cause of anxiety disorders, bipolar disorders, sleep disorders, depression, and occasionally, it can result in suicide and self-harm. As a result, recognition and prompt treatment of people experiencing mental health illness is an extremely difficult endeavour but it has paramount importance. The whole purpose of research is to propose deep learning-based model that primarily incorporates audio and textual aspects of patient answers to diagnose depression using the Extended Distress Analysis Corpus Wizard of OZ (EDAIC-WOZ) dataset. The inclination towards this work is to help psychologists and people with depressive symptoms with an accurate detection using artificial intelligence methods so that quick and timely remedial steps can be taken. The proposed model classifies the patients into two groups based on depression level such as the group of people with depressive disorder and others who doesn’t have such disorder by making use of speech spectral and temporal features and textual emotional analysis. This model uses the late fusion ensemble method implemented on the resultant of Audio CNN model and textual Bi-LSTM model. By using optimized methods for pulling out the textual and audio features, this work ensures that relevant characteristics of audio and text to be considered for the study. The conclusion of the research derived from the data set demonstrate the effectiveness of the model by complimenting the conventional diagnostic methods in screening the depression. The novel method is considered to be efficient and scalable solution to monitor the mental health based on depression level. Results reveal that the proposed model outperforms the other models with F1 score 0.98.