Correlation between frontal facial thermal pattern and affective states of autism spectrum disorder children

Abstract

The prevalence of Autism Spectrum Disorder (ASD) among children in the United States was reported at 1 in 59 for children aged 8 years. In Malaysia, while official statistics are limited, a preliminary investigation carried out by the Ministry of Health Malaysia, examining children aged 18 to 26 months, revealed a prevalence rate of 1.6 per 1000 children. Numerous cases remain undiagnosed, yet the escalating number of autism cases observed by healthcare professionals in paediatric centres strongly indicates a potentially higher prevalence rate in Malaysia. Children with ASD encounter difficulties in expressing affective states, particularly due to a deficit in socio-emotional communication skills. Understanding their affective states is crucial, yet conventional assessment methods, such as EEG and ECG, which often involve the use of patches, can be invasive and may lead to emotional distress. These methods disrupt natural behaviours, leading to inaccurate representations of their affective states. Recognizing the need for a more effective approach, the research proposes non-invasive methods to assess the affective states of children with ASD. It introduces a novel framework for modelling of affective states by using Convolutional Neural Network (CNN) classifier. The research recruited 56 children as the subject, comprising 28 ASD children aged between five and nine years (M = 6.43, SD = 1.2), and an additional 28 typically developing (TD) children (M = 5.65, SD = 2.2) serving as the control group. The investigation focused on the frontal facial thermal imaging of ASD children in response to specially developed video stimuli representing five primary affective states. To ensure the accuracy and impartiality of the assessment, the stimuli were verified by expert blind coders through questionnaires. These questionnaires captured the subjects’ responses to each video stimulus, evaluating valence and arousal levels. The thermal imaging data of children with ASD exhibited unique patterns associated with cutaneous blood flow under the skin regulated by the Autonomic Nervous System (ANS). These patterns were associated with the five basic affective states when the children were exposed to the video stimuli. The research leveraged GLCM, wavelet coefficients, and thermal intensity values from specific regions of interest (ROI) in the facial image as input features for the CNN model. The model enabled real-time computation of affective state outputs, facilitating quantifiable correlations between temperature patterns and affective states. Statistical analysis evaluated these correlations in forms of valence and arousal values. The responses were then mapped onto Kollias's 2-D Circumplex Model of Affect to validate the affective state model. The proposed model is capable of classifying the affective states with high accuracy of 94.10% for TD, 89.60% for ASD, and precision of 95.76% for TD, 91.66% for ASD. The validity of the approach was further confirmed using the CK+ and Rusli et al. frontal facial databases, demonstrating notable performance with an accuracy of 91.81% and precision of 94.54%. These findings reveal the potential of using non-invasive and less intrusive method through thermal imaging with advanced machine learning techniques in assessing real-time affective states in autistic children. It facilitates a more effective diagnostic and early intervention therapy.