Deperssion Detection in Naturalistic Environmental Condition

Abstract

In light of depression's massive and growing burden on modern society, researchers are investigating ways to detect depression early using non-invasive, automated, scalable, and non-invasive methods. In naturalistic environments, though, speech-based methods are still needed for effective articulatory information capture. The intermodality and intramodality affinities are learned by integrating information from audio, video, and text modalities in a multimodal depression prediction model this article. As a result of The most important components of each modality in the multilevel attention are selected to be used for decision-making to reinforce the overall learning. Our study aims to build various regression models using inputs from audio, video, and text landmark duration and historic n-gram features on the DAIC _WOZ and SH2 datasets both performed well, either separately or in combination.