Personalized breathing recommendations for anxiety relief: a data-driven exploratory approach focusing on feasibility

Breathing exercises are widely used to alleviate anxiety, but most interventions use fixed templates of inhalation, exhalation, and repetition. This study preliminarily explores the feasibility of data-driven personalization of breathing recommendations for general individuals. We hypothesized that based on the characteristics of individual users, a data-driven method could give a feasible breathing intervention suggestion without the need of specialized hardware. A multilayer perceptron was trained on questionnaire responses to predict the inhalation, exhalation, and repetition times. The model stability was assessed using several independently trained models and exhibited high consistency with coefficients of variation below 10% across all outputs. A single-blind randomized study (N = 820), which compared the personalized recommendations to a fixed template that is known to effectively reduce anxiety, measured the outcomes in terms of anxiety reduction and user-reported satisfaction. The main analysis and subgroup analysis both detected no statistically significant differences in effectiveness between model recommendation and fixed template after applying multiple comparison corrections. However, adults aged 18-30 years with the medium anxiety level showed a moderate exploratory preference for the fixed template. These findings indicate that individual characteristics may influence the preference of breathing interventions. Overall, the results highlight the feasibility of datadriven personalization in non-medical anxiety interventions. The outcomes can still be enhanced by various methods such as collecting more training data, fine-tuning on the model, or choosing advanced machine learning or offline reinforcement learning approaches.