LSHMMformer: An Intelligent detection model of depression based on multi-modal fusion

Depression is a serious mental health issue, which has become an undeniable threat to mental well-being. Against this background, research on automatic depression detection based on deep learning (DL) can help reduce the cost of depression diagnosis and provide data support for early intervention and the formulation of personalized treatment plans for depression. Current depression detection methods face challenges in efficiency and accuracy. This study aims to develop an intelligent depression detection model to improve prediction performance and enhance detection efficiency.

An intelligent depression detection LSHMMformer was proposed, integrating a novel Locality-Sensitive Hashing Attention mechanism (LSHAttention) and a multi-modal fusion module. The model replaces traditional LSH's fixed random projection matrix with a learnable projection matrix. Multi-modal features are deeply fused using stacked Transformer Decoder layers to comprehensively capture cross-modal correlations between audio and text modalities.

This model was tested on the publicly available depression dataset EATD-Corpus. The mean absolute error of the prediction results was 5.66, and the root mean square error was 8.68.

In comparison with existing methodologies such as GRU and BiLSTM, the proposed LSHcoder module demonstrates superior performance in terms of convergence speed, training stability, and final predictive accuracy, while further enhancing efficacy in multimodal fusion scenarios.

The novel model LSHMMformer significantly improves depression detection accuracy. This model facilitates software development for personalized depression screening, reduces diagnostic barriers in clinical practice, and provides reliable decision support for early intervention.