Thermal imaging-based classifier for affective states of autism spectrum disorder (ASD) children

Abstract

Children with autism spectrum disorder (ASD) are identified as a group of people with social and emotional difficulties. Most of them face challenges in giving the appropriate social response through facial expression and speech. Since emotion is the key to effective social interaction, it needs to understand the correct expressions of emotion and confessions. Emotion is a type of emotional state and can be detected through behavioural interaction, physical interaction, and physiological signals. In general, recognizing emotional states from physical interaction such as facial expression and speech to children with autism is often unexpected. Consequently, an alternative method has been suggested to identify emotional states through physiological signals. Although considered unobtrusive, most current methods require sensor to be patched on to the skin body to measure the signals. This is likely to cause discomfort for children and hide their true emotions. The study suggests the use of thermal imaging technique as a passive method for analyzing physiological signals associated with affective states unobtrusively. The study hypothesizes that the effect of skin temperature changes as a result of pulsating blood flow in the blood vessels in the anterior facial region can be correlated to different affective states for ASD children. As such, a structured experiment was designed to measure the thermal data resulting from the expression of the different emotional state induced using a set of stimuli. To determine vascular regions, the thermal distribution of the face image was analyzed using a grey level occurrence matrix (GLCM). Then, a wavelet-based technique was deployed for patterns detection in time series to spot changes in emotional states throughout stimuli. A set of thermal cues were extracted and statistically analyzed before being fed into the k-Nearest Neighbor (k-NN) classifier to identify the emotional state. In the study, the affective state classification model for typically developing (TD) children aged between 5 to 9 years old was used as a baseline to form an ASD classifier. The results from the classifier showed the efficacy of the technique and accorded good performance of classification accuracy at 88% in identifying the affective states of autistic children. The inter-rater analysis has been done to find the agreement between the classifier’s output and the manual techniques used to detect the affective states in ASD children. The results posed a challenge for therapists to determine the states of ASD children manually through visual observation due to poor expression of contextual emotional states as compared to TD children.