A Novel Multimodal Method for Depression Identification

Abstract

Depression is one of the most prominent mental health issues, characterized by a depressed low mood and an absence of enthusiasm in activities. In terms of early detection, accurate diagnosis, and effective treatment, doctors face a serious challenge from depression, which is a serious global health issue. For patients with this mental disease to receive prompt medical attention and improve their general well-being, early identification is essential. For the purpose of detecting various psychological illnesses including depression, anxiety, and post-traumatic stress disorder, medical audio consultations along with survey responses have been used. A depressed individual displays a range of subtle signs that may be more easily identified by combining the results of multiple modalities. Multimodality involves extracting maximum information from data by using multiple modes, so that the deep learning model can be trained efficiently to give better results. Given that each modality functions differently, combining various modalities is not easy, and each origin of a modality takes on a different form. It is clear from the literature that is currently significant in the area that, combining the modalities yields positive outcomes. A trustworthy approach to identify depression is thus urgently needed because it continues to be a problem for many individuals in today’s society. In this work, textual and audio features are incorporated related to the identification of depression, and a novel multimodal approach using an optimized Bi-directional Long Short -Term Memory model that recognizes premature depression is suggested for medical intervention before it develops further.

A literature review on multimodal deep learning models for detecting mental disorders in conversational data: Pre-transformer and transformer-based approaches

“…BiLSTM model hyperparameters include the learning rate, the number of epochs, the optimization method, the loss function, and the activation function for the dense layer. The 74.3% accuracy and 5.14% error rate attained by the multimodal technique are superior to individual modalities, as demonstrated in the study of Rahul Singhal, et al [16].…”

“…The structure of this system is shown in Figure 3 [15]. Singhal et al [16] proposed an alternative method for combining audio and text data using a Bidirectional Long Short-Term Memory (BiLSTM) model. The BiLSTM model is a variant of recurrent neural networks that processes input sequences both forward and backward.…”

Taking All the Factors We Need: A Multimodal Depression Classification With Uncertainty Approximation

“…Recent research has presented po-tential remedies to these obstacles. Some of these remedies involve the utilization of BiGRU, BiLSTM [36], [37], and Hierarchical Attention Network (HAN) [28] architectures for text analysis. Other approaches involve the application of GPT2-medium language models to generate task-oriented embeddings [26].…”