Analysing neural network models for detecting panic attacks with uncertainty analysis

Abstract

In our society and around the world a lot of people suffer from panic attacks. These panic attacks can be mild or very intense physical stimulations that may incapacitate an individual at the spot when the panic attack occurs. The problem in this case is, if the person suffers from a panic attack outside their house and loses control over themselves, they might be subjected to external environmental hazard such as getting into a car accident, etc. Therefore, if we can effectively track and detect whether a person had a panic attack via their spatiotemporal and biometric data, steps can be taken to help them recover from the panic attack or send help to them, as quickly as possible. Keeping this in our mind, in this study we analysed the performance of different neural network models and techniques to detect panic attacks of individuals from their spatiotemporal and biometric data. Since detection of panic attacks is an emergency use-case, model reliability is essential. To ensure model reliability, we also represented the uncertainty analysis of these neural network models using Monte Carlo Dropout. During our study, we found that among all the models that were used, GRU (Gated Recurrent Unit) had the highest accuracy of 95.56%, and GRU also had one of the least amount of uncertainty. However, the ensemble model had the least amount of uncertainty among all the models that were used.