Application of Machine Learning for Panic Attack Detection using Health Wearable Sensors

Nowadays, with people becoming more conscious of their physiological and psychological conditions, wearable health devices are becoming increasingly ubiquitous. Consequently, different types of edge or cloud computing models are proposed to provide users with other monitoring and user conditions. While many have proposed predicting stress, this study presents the framework to label and predict panic attacks from wearable sensors, such as electrodermal activity, heart rate, and skin temperature, which are significant indicators of symptoms like rapid heart rate and sweating. The framework for labeling panic attacks is based on an open-source nurses dataset with over 11 million data. The dataset also provides information on users' stress conditions, labeled no stress, low stress, and high stress, which are exploited to provide the labeling of users having the attack. The dataset was used to test different machine-learning models. The highest test accuracy was achieved by the Medium and Fine Tree models with 99.96%, followed by Cosine K-Nearest Neighbour (KNN) (88.44%) and Coarse KNN (87.81%). The study proves that multiple models can be used to make the prediction, but expert opinions should be obtained to help with a more reliable automated classification of users' conditions using health wearables. As the selected sensors are common in modern health wearables and sold commercially, the study is believed to benefit millions of people who are exposed to the disorder.