Deep learning-based anxiety quantification assessment and emotional regulation recommendation generation model

Accurately managing emotions and quantifying anxiety continue to be important challenges in affective computing and digital mental health, especially where continuous, objective, and personalized evaluation goes beyond self-reported anxiety. Current deep learning systems mainly have coarse emotion classification ability and have no built-in and interpretable regulation mechanisms. In order to overcome these limitations, the Deep Learning Based Anxiety Quantification and Emotional Regulation Recommendation Generation Model (DLAQ-ERRG) is proposed in this study. It comprises three key components: (i) a Neural Stochastic Differential Attention Encoder (NSDAE) for capturing the continuous-time affective dynamics and temporal uncertainty from multimodal wearable signals; (ii) a Hyperbolic Contrastive Affective Manifold Learner (HCAML) for learning a non-Euclidean space to embed latent emotional states while preserving the hierarchical structure of affective severity and boosting inter-subject separability; and (iii) a Hyperbolic Neural Symbolic Policy Generator (ER-NSPG) for generating constraint-consistent and uncertainty-aware emotional regulation recommendations. Experimental results under strict Leave-One-Subject-Out validation on the WESAD dataset show an improvement of 11.85% compared to the best baseline (CCC = 0.734) and the Mean Absolute Error (MAE) of 0.384 and the Root Mean Square Error (RMSE) of 0.083, respectively. The overall anxiety prediction accuracy (AIPA) for affect-state classification is 87.3% for the model. Structured severity organization is confirmed by representation quality metrics: an Anxiety Severity Separation Score (ASS) of 0.726 and a Hyperbolic Clustering Quality Index (HCQI) of 0.681. The recommendation evaluation results in an Action Appropriateness Score (AAS) of 0.776, Recommendation Smoothness Score (RSS) of 0.764 and Emotion Regulation Diversity Score (RDS) of 0.701, all representing stable and tailored generation of interventions. The results illustrate the benefits of uncertainty-aware continuous estimation of anxiety-related arousal, improved structural representation, and interpretable support in regulating arousal within a single, uncertainty-aware framework.