Panic Attack Prediction Using Wearable Devices and Machine Learning: Development and Cohort Study (Preprint)

Abstract

BACKGROUND A panic attack (PA) is an intense form of anxiety accompanied by multiple somatic presentations, leading to frequent emergency department visits and repetitive medical examinations before a formal diagnosis. Designing a PA prediction model could help prevent iatrogenic harm to patients and facilitate more personalized treatment. OBJECTIVE This study aimed to provide a seven-day PA prediction model and determine the relationship between physiological factors, anxiety and depressive factors, and air quality index. METHODS We enrolled 59 participants with PD (DMS-5 and MINI interview). Participants used smartwatches (Garmin vivosmart 4) and mobile applications to collect their sleep, heart rate, activity level, anxiety, and depression scores (BDI, BAI, STAI-S, STAI-T, and PDSS-SR) in their real life for a duration of one year. We also included air quality indexes from open data. To analyze these data, our team used six machine learning methods: random forests, decision trees, LDA, AdaBoost, XgBoost, and regularized greedy forests. RESULTS For seven-day PA prediction, the random forest produced the best prediction rate. The model achieved an accuracy of 97.5% on the training set. Overall, the accuracy of the testing set was 67.4–81.3% for different machine learning algorithms. The most critical variables in the model were questionnaire and physiological features such as BAI, BDI, STAI, MINI, average heart rate, resting heart rate, and deep sleep duration. CONCLUSIONS It is possible to predict panic attacks using a combination of data from questionnaires, and physiological and environmental data. Prediction accuracy was 97.5% on the training data and 81.3% on the testing data.