Novel Depression Classification Framework Using Optimal Feature Integration with Hybrid Convolution (1D/3D) Based Adaptive Residual DenseNet

Abstract

This study introduces a novel depression classification method by incorporating multimodal data to reduce this limitation and improve the accuracy of depression diagnosis. In the beginning, the multimodal data such as speech signal, Electroencephalographic (EEG), and text data is obtained from the benchmark datasets. These acquired text data are subjected to the text pre-processing phase, where the stemming, character removal, punctuation and stop word removal operations are performed. After that, the resultant text is given to the Bidirectional Encoder Representations from Transformer (BERT), and the extracted features are considered Feature 1. From the EEG signals, feature 2 is attained from the wave features. Accordingly, feature 3 is attained from the linear and non-linear features. Finally, from the speech signals, the spectral feature is extracted and is considered Feature 4. Further, the extracted four features are optimally fused by using the proposed Modified Random Value of Osprey Optimization Algorithm (MRVOO). Subsequently, the optimally fused features and the video frames are subjected to the Hybrid 1D and 3D Convolution-based Adaptive Residual DenseNet (HCARDNet) for depression classification. Here, the network parameters are optimally determined by the same MRVOO. The performance is examined via distinct metrics and it outperforms with the better classification rather than baseline approaches.