Anticipating Panic Attacks Using Fitbit: A Pilot Study Among College Students in Lebanon

Abstract

With the increasing availability and accuracy of information provided from smart wearables, this study explores the feasibility of using wearables data to predict and diagnose panic attacks among AUB students in Lebanon. Fourteen participants wore Fitbit Inspire 3 devices over a 2-month period to collect physiological features such as heart rate, heart rate variability, physical activity, and sleep data. Participants submitted daily logs if they experienced any panic attack events. Data was processed into 15-minute intervals and analyzed using machine learning models to classify and predict panic attack occurrence for the same day (PA Today) and the following day (PA Tomorrow). Statistical analysis revealed significant differences in heart rate and sleep data on days leading up to panic attacks, including increased light sleep, restlessness, and elevated heart rate. These features along with others were used to train three machine learning models, including Random Forest, XGBoost, LSTM, and ensemble combinations. Random Forest achieved the highest F1 score for PA Today classification (0.421), while XGBoost outperformed other models in forecasting next-day panic risk (F1 = 0.400). Ensemble and LSTM models showed promise but did not outperform individual classifiers. The results demonstrate the potential for non-intrusive, wearable devices to support early warning and diagnosis of panic attacks. Future research should focus on developing intervention methods based on panic attack risk prediction, refining ensemble and time series methods, and developing an adaptive system for broader populations.