| dc.description.abstract |
The diagnostic process for Autism Spectrum Disorder (ASD) is complex,requiring extensive expertise to integrate information from diverse sourcessuch as parental reports and clinical observations. However, limited accessto diagnostic facilities, especially in suburban areas, poses a significant challenge,often necessitating multiple expert consultations before an accuratediagnosis can be made. Given the shortened neuroplasticity period experiencedby children with ASD (CwA), early intervention is crucial for enhancingtheir social and communication skills. These challenges have promptedthe research community to explore technology-based solutions to meet theneeds of individuals with autism, caregivers, and professionals, with a focuson achieving reliable and objective diagnoses.Artificial Intelligence (AI)-based behaviour analysis has emerged as apromising approach for identifying characteristic behaviours associated withASD diagnosis. With this in mind, the thesis aims to develop an AI-assistedassessment system capable of identifying these traits in children with autism.Moreover, recognizing the preference of children with ASD for technologyassistedtools in therapeutic interventions, robot assistants were employed tointeract with children during diagnostic procedures, aiming to create engagingand comfortable interactions for the children undergoing assessment. Theoverarching goal is to develop an assessment system that is context-aware, robust in adhering to diagnostic standards, and transparent in its decisionmakingprocesses, thereby addressing current gaps and challenges in existingintervention approaches.To accomplish these objectives, we initially assessed the feasibility of deployingthe system in resource-constrained environments, particularly in theGlobal South, with a specific focus on India. Our investigation began byexamining the response behaviour of children of Indian ethnicity, both withand without autism spectrum disorder, towards robot-assisted interventionsfor diagnosis. Through a study targeting children aged 3 to 6 years, weobserved that the children effectively followed the robot’s instructions duringdirective tasks and successfully completed them. Our analysis revealedthe acceptance and benefits of employing robotic assistants as facilitators invarious domains, including education, cognitive therapies, and healthcare.Subsequently, we conducted a study to explore the perceptions of specialeducators regarding the use of social robots in robot-assisted interventionsfor diagnosis. Employing a mixed-methods approach involving interviews,workshops, and a panel discussion with 25 educators in India, we uncoveredboth challenges and opportunities associated with integrating social robotsinto autism interventions. While special educators expressed concerns abouttheir functional capacity and apprehensions regarding potential redundancyresulting from the substitution of human efforts by social robots, they alsoacknowledged the importance of technological innovation in reshaping andenhancing their roles in autism therapy. Despite initial scepticism, profes sionals identified various strategies for effectively incorporating social robotsinto intervention programs. We further elucidate the implications of thesefindings for the development of context-aware solutions and policy-level initiativesessential for resource-constrained settings.Having established the feasibility of robot-mediated interventions for childrenwith autism, our focus shifted to developing interpretable machinelearningframeworks capable of identifying speech and facial expression behavioursessential for ASD assessment and diagnosis. Our initial endeavourinvolved the creation of a multi-source transfer learning approach to capturefacial emotional attributes in children of Indian ethnicity, thereby aiding diagnosticdecisions. Through optimization techniques aimed at enhancing multivariatecorrelation among source tasks, we achieved notable improvementsin facial emotion recognition accuracy compared to existing methodologies.Subsequently, we turned our attention to identifying speech behaviouralcharacteristics critical for ASD diagnosis, emphasizing language impairmentsas key markers. Through automated methods for speech behaviour extractionand comprehensive speech evaluations, we characterized linguistic traitsin children with autism speaking Hindi. Our extensive analysis encompasseda diverse set of acoustic and linguistic speech attributes, including lexical,syntactic, semantic, and pragmatic elements. Notably, our investigation alsoaddressed the influence of linguistic diversity on speech-based ASD assessment,examining speech behaviours in both English and Hindi-speaking children.The results of our study, involving data from 76 English-speaking children and 33 Hindi-speaking children, demonstrated the reliability of automaticallyextracted acoustic and linguistic features as predictors of ASD,achieving an impressive macro F1-score of approximately 91.30%.Concluding our thesis, we validated the AI-assisted system with a robotmediator for the diagnostic assessment of children with ASD through a pilotstudy involving 21 participants. This involved 12 typically developing (TD)children and 9 CwA. The results affirmed the feasibility and efficacy of therobot-mediated AI system in diagnosing ASD in Indian children. In sum,our thesis provides an initial exploration into the potential of utilizing robotsin autism care in India, supported by AI-assisted multimodal behaviouralanalysis, paving the way for further advancements in this field. |
en_US |