What question did this study set out to answer?

July 13, 2026Open Access

Applications of deep learning in the identification and classification of mental health status

XWXiaolei WangBeijing Polytechnic SYShuo YangShanghai Public Security Bureau XFXiaohui FengBeijing Polytechnic

Key Points

The research aims to enhance the identification and classification of mental health status through advanced deep learning techniques.
Developed a Chinese-RoBERTa-based multi-head attention fusion sentiment analysis framework (CR-MHA FSA).
Combined hierarchical emotional knowledge with multimodal features and utilized BiLSTM for feature extraction.
Applied focal loss to improve learning from imbalanced emotional categories.
BiLSTM achieved an average F1 score of 0.817 on the EMO-DB task, outperforming VGG (0.745) and LSTM (0.802).
F1 scores for strong positive stimuli during question-answering stages reached 0.5938 and 0.5518.
Pearson similarity between behavioral entropy and depression trends exceeded 72.2%, indicating a strong correlation.

Abstract

This research aimed to address the requirements of the identification and classification of mental health status. It tackled the challenges of traditional scale assessments, such as difficulty in continuous monitoring and high subjectivity, as well as the historical influence decay, sequential dependence, and imbalanced category distribution of emotional expression in psychological interview scenarios. A Chinese-RoBERTa-based multi-head attention fusion sentiment analysis framework (CR-MHA FSA) framework integrating hierarchical emotional knowledge and multimodal features was developed to solve these issues. The model used Chinese-Roberta-Whole Word Masking (WWM)-ext to obtain contextual semantic representations and combined Bidirectional Long Short-Term Memory (BiLSTM) to extract sequential emotional features. A tree-structured hierarchical label space was established for the characterization of the granularity of emotional semantics and label dependence and multi-head self-attention was applied to obtain feature fusion. To better reflect emotional memory mechanism in interviews, this research fitted forgetting curve to emotional word weight allocation function, introducing the decay of historical emotional words over time. Simultaneously, focal loss was applied to decrease the weight of a large number of simple samples to enhance few-sample learning. Experimental findings showed that on Berlin Database of Emotional Speech(EMO-DB) source task, BiLSTM achieved an average F1 score of 0.817, outperforming Visual Geometry Group(VGG) (0.745) and Long Short-Term Memory (LSTM) (0.802). In the anomaly recognition induced by emotional stimuli, the F1 scores for strong positive stimuli in the question-answering text reading stages were 0.5938 and 0.5518, respectively. Pearson similarity between the behavioral entropy depression trend and total score trend was > 72.2%, reaching a maximum of 85%, with F1 scores for "fear" and "surprise" increasing by 2.41 and 1.37 percentage points, respectively. This research implemented a multimodal mental health detection and assessment system as well as early warning process to form a closed loop "identification—grading—reporting/early warning".

KI fragen

Bookmark

View Full Paper