Upset or Collapse Detection System for ASD Children Using Smart Watch with Machine Learning Algorithm

Abstract

ASD is characterised by severe and violent behavioural issues that are referred to as "meltdowns (upset) or tantrums (collapse)" and can include aggression, hyperactivity, intolerance, unpredictability and self-injury. This research work intends to develop and implement a non-invasive real-time Upset or Collapse Detection System (UCDS) for people with ASD. With a certain model of smart watch, the non-invasive biological indications such as Pulse Rate (PR), Skin Temperature (ST), and Galvanic Skin Reaction (GSR) can be artificially captured.  In order to create the UCDS, deep learning algorithms like CNN, LSTM, and the hybrid of CNN-LSTM are given the physiological signals that are captured to a server. The deep learning algorithm could recognise aberrant upset or collapse states from real-time physiological signs after being trained.  Deep learning algorithms including CNN, LSTM, and CNN-LSTM are used to train and test the proposed UCDS system, and it is discovered that hybrid CNN-LSTM beat them all with an average training and testing accuracy of 96% and a low mean absolute error (MAE) of 0.10 for training and 0.04 for testing.  Furthermore, the suggested UCDS system is supported by 93% of the ASD caretakers.