Adaptive Robotic Techniques in Children with Autism: Strategies for Utilizing Physiological Data to Optimize Engagement during Computer-Based Interactions

Background:

While it is well documented that Autism Spectrum Disorders (ASD) are characterized by impairments in social interaction and communication, questions remain about the neurobiological mechanisms and physiological processes underlying these core-defining behaviors. Recent technological advances have shown initial promise in identifying physiological components associated with social information processing in specific contexts. Despite advances, researchers have not yet successfully developed systems for mapping physiological response systems that can be flexibly applied in real time and real world circumstances.

Objectives:

In the present study, we attempted to develop and apply an adaptive response technology that identified and flexibly modified components of computer-based interactions to optimize engagement in a sample of children with ASD.

Methods:

Six children with autism (ages 13-16 years) and a PPVT-III score of 80 or above participated in six sessions of computer tasks – three solving anagrams and three playing Pong – followed by two sessions of robot-based basketball. A child was involved in the computer tasks while his/her physiological data (i.e., cardiovascular (ECG), electrodermal (EDA), and electromyographic (EMG) signals) were acquired via wearable biofeedback sensors. Changes in task difficulty were designed to evoke varying intensities of three target affective states: liking, engagement, and anxiety. An observing therapist and parent of the participant provided subjective reports during the tasks. Each task session was subdivided into a series of discrete trials, termed epochs. After each epoch subjective reports were collected from the child, his/her parent, and the observing therapist on the child's level (high/low) of the three target affective states during the previous epoch. Physiological data and subjective reports collected during the human-computer interaction tasks were used to train an individualized psychophysiological model. A machine-learning technique, Support Vector Machines, was used to build psychophysiological models, which map between the physiological features and ratings of engagement/affective state. We then evaluated models of interaction where the robot selected its behavior randomly and where the robot autonomously selected its behavior based on the child's current physiological information and predictions derived from their own unique individual psychophysiological model.

Results:

From the closed-loop tasks, we found a robot could maintain and possibly increase subjective ratings of liking during autonomous, real-time interaction with a child with ASD relative to random selection. Prediction accuracies from the psychophysiological models improved with additional physiological information with EMG signals being less discriminatory than ECG and EDA signals.

Conclusions:

Pilot data support the use of psychophysiological modeling as a viable technique for further exploration and incorporation into intervention and interactive experiences for children with ASD. Work incorporating adaptive physiological monitoring into virtual reality platforms aimed at exploring specific social communication and sensory vulnerabilities of children with ASD is underway.