AI Model Reads Anxiety From Wearable Signals and Suggests Ways to Calm Down

Anxiety is one of the most common mental health complaints in the world, yet the tools we use to measure it remain stubbornly old-fashioned. Questionnaires filled out in a clinic capture a snapshot of how someone feels at a single moment, filtered through memory and a natural reluctance to admit distress. A new study published in Discover Artificial Intelligence proposes something far more ambitious: a deep learning system that reads anxiety-related physiological arousal continuously from wearable biosignals, expresses it as a number on a continuous scale, and then generates personalized, interpretable recommendations for emotional regulation. The framework, called DLAQ-ERRG, was developed by Jing Zhai of Anshan Normal University and Qiang Wan of Hanseo University and South China University of Technology.

The researchers argue that most existing approaches to anxiety detection suffer from a fundamental conceptual flaw: they treat anxiety as a discrete category, a yes-or-no label, rather than as a mental state that ebbs, flows, and varies in intensity over time. Convolutional and recurrent neural networks, the workhorses of physiological signal analysis, also struggle to incorporate uncertainty, inter-subject variability, and the hidden hierarchy of emotional severity. And crucially, most systems stop at prediction. They tell you that someone seems anxious, but offer no guidance on what to do about it. DLAQ-ERRG was designed to close that gap with a single end-to-end pipeline that moves from raw sensor data to actionable, explainable support.

The architecture rests on three interlocking components. The first, a Neural Stochastic Differential Attention Encoder (NSDAE), models the latent anxiety state as a continuous-time stochastic process governed by a stochastic differential equation. Instead of chopping biosignals into static windows and treating each one independently, the encoder learns a drift function, parameterized by an attention mechanism that dynamically weights the contribution of each physiological modality, and a diffusion term that captures random fluctuations, sensor noise, and individual differences. Amortized variational inference with a Gaussian posterior allows the model to output not just a point estimate of arousal but a calibrated confidence interval around it, distinguishing sudden emotional reactions from enduring patterns.

The second component, a Hyperbolic Contrastive Affective Manifold Learner (HCAML), addresses a subtler representational problem. Mild, moderate, and severe states of arousal form a hierarchical, tree-like structure, and flat Euclidean latent spaces handle such geometry poorly. HCAML projects the learned representations onto a Poincaré ball with fixed curvature of minus 1.0, where distances expand exponentially toward the boundary of the manifold. In this curved space, low-arousal states cluster near the center while higher-severity states spread toward the periphery, and a radial ordering regularization term enforces that greater severity corresponds to greater distance from the origin. A contrastive loss with an InfoNCE formulation, using a temperature of 0.07, pulls temporally adjacent windows from the same subject and condition together while pushing apart samples from different subjects and conditions.

The third component, an Entropy-Regularized Neuro-Symbolic Policy Generator (ER-NSPG), turns quantified arousal into decisions. The action space is deliberately small and human-readable: paced diaphragmatic breathing at six to eight breaths per minute, guided mindfulness grounding, cognitive reappraisal prompts, a brief physical pause or stretch, a social support suggestion, or passive monitoring. Symbolic rules derived from cognitive-behavioral intervention logic constrain which actions are admissible at a given arousal level. When predicted intensity exceeds a high threshold, only physiological down-regulation strategies such as paced breathing and grounding are permitted; moderate states allow cognitive reappraisal; low-intensity states default to monitoring. A reinforcement learning policy trained with proximal policy optimization, entropy regularization, and temporal smoothness penalties selects among the admissible actions, preventing both overconfident determinism and erratic switching between interventions.

Training and evaluation relied on the WESAD dataset, a public benchmark of wearable stress and affect detection containing chest- and wrist-worn sensor recordings from 15 laboratory volunteers, including electrocardiography, electrodermal activity, respiration, and blood volume pulse. Because WESAD lacks continuous anxiety annotations, the team constructed a proxy target: a normalized arousal index blending experimentally induced stress activation with min-max-normalized self-reported stress ratings, weighted 0.7 and 0.3 respectively. The authors are careful to stress that this index represents anxiety-related physiological arousal in an affective computing sense, not a clinical diagnosis of an anxiety disorder. Validation used a strict Leave-One-Subject-Out protocol, in which each of the 15 subjects was held out in turn, with all preprocessing parameters estimated exclusively from training subjects to eliminate data leakage.

The results are striking. The full framework achieved a Concordance Correlation Coefficient of 0.921 plus or minus 0.028, a Mean Absolute Error of 0.061, and a Root Mean Square Error of 0.083, with classification accuracy of 97.3 percent for the three-class affect task. Compared against the strongest baseline, a Transformer encoder scoring 0.882, the improvement was statistically significant, with a paired t-test yielding p equal to 0.0003 and a large effect size of 1.26. Ablation studies confirmed that each component earns its place: replacing the stochastic encoder with a BiLSTM dropped the CCC to 0.889, while swapping hyperbolic embeddings for Euclidean ones of identical parameter count reduced the Anxiety Severity Separation Score by 6.8 percent. Removing entropy regularization cut recommendation diversity by 11.2 percent, and removing smoothness regularization increased temporal action variance by 14.7 percent.

The recommendation module was evaluated on multiple fronts. It scored 0.776 on the Action Appropriateness Score, 0.764 on the Recommendation Smoothness Score, and 0.701 on the Emotion Regulation Diversity Score. A subset of 120 state-action pairs was independently rated by three domain-informed evaluators, reaching a Cohen’s kappa of 0.81, and the appropriateness score correlated with mean expert ratings at r equal to 0.74. Case studies illustrate the system in action: for one subject in a high-arousal stress phase, predicted intensity of 0.81 triggered three consecutive windows of paced breathing, after which predicted intensity fell to 0.63 within 15 seconds and the uncertainty band narrowed. In a failure case involving motion artifacts, the model’s prediction briefly spiked while ground truth held steady, but the elevated uncertainty flagged the problem and the system reverted to passive monitoring once signal variance dropped.

Calibration, often the Achilles heel of probabilistic models, proved strong here. The Expected Calibration Error was just 0.027 across folds, and the correlation between predicted standard deviation and absolute regression error was 0.61, meaning the model genuinely knows when it is unsure. Discarding the top 10 percent of highest-uncertainty samples reduced RMSE from 0.083 to 0.071, demonstrating that the uncertainty estimates could support reliability-aware deployment. On the computational side, the full model contains roughly 2.31 million parameters and runs a forward pass in 18.6 milliseconds on a GPU or 74.3 milliseconds on an ordinary Intel i7-class CPU, comfortably fast for real-time wearable use. A pruned wrist-only configuration with a reduced latent dimension cuts inference latency to under 50 milliseconds with a negligible accuracy penalty.

The authors are candid about the limits. WESAD contains only 15 subjects recorded under controlled laboratory conditions with an acute, experimentally induced stress protocol, so generalization to chronic anxiety, clinical populations, and real-world environments remains unproven. The symbolic rules are fixed rather than adapted to individual coping styles or cultural context, and the recommendations have not been validated as therapeutic interventions in outcome trials. A small pilot study with 18 participants found high ratings for clarity and appropriateness of the explanations, but noted that overly frequent recommendations hurt perceived usability. Privacy of sensitive physiological data, fairness across demographics, and the need for human oversight all receive explicit attention. Still, as a proof of concept, the work points toward a future in which a smartwatch does not merely count your steps, but quietly tracks your autonomic state, tells you how it knows, and suggests a breathing exercise at exactly the moment it might help.

Subject of Research: Deep learning-based continuous quantification of anxiety-related physiological arousal and generation of personalized emotional regulation recommendations from multimodal wearable signals

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

Article References: Zhai, J., & Wan, Q. (2026). Deep learning-based anxiety quantification assessment and emotional regulation recommendation generation model. Discover Artificial Intelligence, 6(1), Article 1357. https://doi.org/10.1007/s44163-026-01886-w

Image Credits: AI Generated

DOI: 10.1007/s44163-026-01886-w

Keywords: anxiety quantification, affective computing, deep learning, wearable sensors, stochastic differential equations, hyperbolic representation learning, neuro-symbolic reasoning, emotion regulation, uncertainty-aware modeling, WESAD dataset, reinforcement learning, digital mental health