Heart Rate Predictive Value Using Wearable Sensors in Social Robotics Conversations to Help Children with Autism

The problematic behavior in children suffering from autism, especially in the skills of social and communication, is a problem for family, therapists, caretakers, and also in the treatment process. The work presented in this paper investigates the potential for different machine learning algorithms, applied to data collected through children’s wearable sensors while interacting with toys and social robots. Data were collected through the wearable devices and video annotations from the sessions for identification of problematic behaviors. The extracted time features were then integrated with four machine learning algorithms for detection of problematic behavior: the mean, standard deviation, min, and max. The research also takes into consideration the variations in HRV. Among the included machine learning techniques, the XGBoost algorithm has shown a higher rate of better performance and outperformance in detecting problematic behaviors, with $91 \%$ accuracy. Physiological features outperform kinetic ones in the detection of problematic behaviors, while the predicting factor of one significant thing is in detecting heart rate accuracy. Difficult behavior in this study is proven to be associated with one particular HRV measure, RMSSD. This paper also highlights the importance of development of the instruments and the strategies necessary for the detection of problematic behaviors in autistic children during the social robot-aided session.