Multimodal Representation of Depression and Anxiety Severity Prediction Using Adaptive Capsule Network with Electric Eel Foraging Optimization

Abstract

The assessment of a person's mental state using data-driven models to gauge the severity of their symptoms is known as depression and anxiety severity prediction. This process aids in diagnosis, treatment planning, and therapy progress tracking. Predicting depression and anxiety severity from audio and video data is essential, and various techniques have been implemented to achieve this. However, the existing methods have lack of accuracy, precision and high error rate. To overcome the aforementioned problem, Adaptive Capsule Network (ACN) with Electric eel Foraging Optimization (ACN-EeFO) is proposed for accurately predicting depression and anxiety severity from audio, video data. In this input image is taken from two datasets such as AVEC2013 and AVEC2014 datasets. Then the video and audio data are extracted using Multi-Axis Vision Transformer (MaxViT) and Block-Based Haar Wavelet Transform (B-BHWT). Following that, the extracted audio and video data are fused using Efficient Long-range Attention Network (EL-AN). Then the classification is done by using Adaptive Capsule Network  and optimized with Electric eel Foraging Optimization (ACN-EeFO)  for classify the different stages of depression levels using  BDI-II score. The introduced system is executed in python. The efficiency of the proposed ACN-EeFO is analyzed using 2 datasets and attains 99.82% accuracy, 99.23% F1-Score and attains better results compared with the existing methods. In the future, instead of merging the audio and video features at the end, it will attempt to find out if doing so enhances prediction performance. Additionally, scientists plan to utilize this framework for additional tasks including the prediction of various diseases.