Multimodal emotion recognition-driven personalized digital therapeutics for anxiety management

Anxiety disorders are among the most widespread mental health challenges, yet conventional treatments face barriers of accessibility, cost, and reliance on subjective measures. Digital therapeutics offer scalable solutions, but current systems lack real-time emotion monitoring and adaptive personalization. To address this gap, this study proposes a multimodal emotion recognition-driven framework for personalized anxiety management. The framework fuses electroencephalography, heart rate variability, facial expression, and speech features via cross-modal attention, and employs a reinforcement learning–based decision engine to dynamically select interventions such as breathing exercises, mindfulness, or cognitive reframing. Adaptive feedback further tailors interventions to user responses. Experiments on DEAP and WESAD datasets showed superior performance over unimodal and traditional fusion baselines, with accuracies of 86.2% and 84.7% and AUROCs of 0.91 and 0.89. Anxiety reduction analysis demonstrated up to 24% improvement in State-Trait Anxiety Inventory scores. The study advances affective computing by linking multimodal sensing with adaptive therapeutic design, and offers a foundation for scalable, interpretable, and clinically relevant digital mental health interventions.