Autism is characterised by a non-verbal communication that differs from that of a typically- developing child. "It differs on several points, such as the difficulty in establishing eye-contact, smiling, pointing to objects or the way they are interested in what surrounds them", explains Nada Kojovic, a researcher in Marie Schaer’s team and first author of the study.
Over a period of three years, the scientists, supported by the Swiss National Centre of Comptence in Research (NCCR) Synapsy, developed this algorithm, which aims to classify videos based solely on the child’s movements when interacting with another person. To do this, they first used a technology named OpenPose, developed at Carnegie Mellon University. This computer vision technology extracts the skeletons of moving people as captured in a video and allows the analysis of gestures by removing all characteristics that could be discriminating (age, sex, setting, etc.), keeping only the relationships of skeletons in space and time. The UNIGE research team then developed their AI algorithm tailored for detecting autism and tested it on 68 typically developing children and 68 children with autism, all under 5 years. “We divided each group into two: the first 34 in each group ‘trained’ our AI to differentiate the non-verbal behaviour of children with and without autism. The others then helped us test its accuracy. We also carried out an assessment on 101 other children”, explains Thomas Maillart, a researcher at the Institute of Information Sciences and a faculty member at the Geneva School Economics and Management (GSEM) and of the University Centre for Informatics (CUI) at UNIGE.
The AI sifted through videos of children playing freely with an adult. The results show that the AI makes accurate autism classification in more than 80% of cases. “This is an excellent result“, enthuses Marie Schaer. "In 10 minutes, we can indeed obtain a first screening accessible to anyone, wherever they live. This would allow parents worried about their young children to obtain an initial automated assessment of the symptoms of autism." In addition, this automated video analysis offers complete anonymity.
It should be noted that this technology does not require any direct intervention on the child. "The installation of movement sensors is time-consuming and sensitive; it can also disturb the children and influence the results. Here, the computer vision-based analysis is non-invasive", emphasises Nada Kojovic. Moreover, as it does not require any specific setup, the algorithm can be used to analyse videos recorded in the past, a clear advantage for research purposes.
The aim of the multidisciplinary team is now to make this AI available to everyone. "We now wish to develop an application that would allow such as analysis with only a 10 minutes video filmed with a smartphone", concludes Thomas Maillart.
MEDICA-tradefair.com; Source: Université de Genève
