Anticipating the Risk of Panic Attacks Using Wearables

Abstract

Panic disorder (PD) is a common, often misdiagnosed condition marked by recurrent panic attacks (PA) that impair daily functioning. Despite available treatment, predicting PA onset remains challenging. This study evaluates a wearable-based framework using Fitbit inspire 3 smart watch data to support PD diagnosis and risk of next-day PA prediction. Thirty-two participants wore the device continuously for 2-months. Deep learning ensembles, using a 7-day input window, demonstrated the strongest performance. The GRU model achieved the highest accuracy (0.902) and AUC (0.926), while the LSTM model showed the highest recall (0.995), indicating strong detection of true cases. SHAP analysis identified GAD scores, restlessness, and sleep stages as key predictors. Nonetheless, heart rate (HR) significantly increased during panic periods (interval: t=7.9, p<0.01; day: t=49.43, p<0.01), with greater elevation observed in males (p=0.003). Light sleep also increased during panic nights (~12.3mins, p=0.032). Finally, interviews conducted revealed awareness of early warning signs, preference for coping strategies such us breathing and movement, and strong interest in discreet, real-time PA prediction. Future work should address key limitations, including small sample size, potential biases and incompliance, while guiding the design and commercialization of a system for management and active mitigation.