Early prediction of Autism Spectrum Disorders through interaction analysis in home videos and explainable artificial intelligence

There is considerable discussion about the advantages and disadvantages of early ASD diagnosis. However, the development of easily understandable and administrable tools for teachers or caregivers in order to identify potentially alarming behaviours (red flags) is usually considered valuable even by scholars who are concerned with very early diagnosis. This study proposes an AI pre-screening tool with the aim of creating an easily administrable tool for non-competent observers useful to identify potentially alarming signs in pre-verbal interactions. The use of these features is evaluated using an explainable artificial intelligence algorithm to assess which of the proposed new interaction characteristics were more effective in classifying individuals with ASD vs. controls. We used a rating scale with three core sections – sensorimotor, behavioural, and emotional – each further divided into four items. By seeing home videos of children doing everyday activities, two experienced observers rated each of these items from 1 (highly typical interaction) to 8 (extremely atypical interaction). Then, a machine learning model based on XGBoost was developed for identifying ASD children. The classification obtained was interpreted through the use of SHAP explanations, obtaining an area under the receiver operating curve of 0.938 and 0.914 for the two observers, respectively. These results demonstrated the significance of early detection of body-related sensorimotor features.