Iranian Exceptional Children Foundation Empowering Exceptional Children 2588-3488 17 2 2026 06 22 Qualitative Research Synthesis of the Application of Deep Learning in Video Analysis of Social Interactions for Early Diagnosis of Autism Spectrum Disorder in Preschool Children Qualitative Research Synthesis of the Application of Deep Learning in Video Analysis of Social Interactions for Early Diagnosis of Autism Spectrum Disorder in Preschool Children 66 42 246854 10.22034/ceciranj.2026.566116.2013 FA Rahim Moradi Associate Professor, Department of Educational Sciences, Faculty of Humanities, University of Arak, Arak, Iran 0000-0001-9401-9631 Mozhgan Ghanat M.A. in Educational Technology, Faculty of Psychology and Educational Sciences, Allameh Tabataba’i University, Tehran, Iran 0009-0006-1274-9467 Journal Article 2025 12 14 Early diagnosis of autism spectrum disorder (ASD) in preschool children requires a precise analysis of social interactions, which is often prone to subjectivity and human error when using traditional methods. Deep learning, with its ability to automatically extract complex patterns from video data, offers a novel opportunity to enhance diagnostic accuracy and objectivity. This research synthesis aimed to systematically analyze the applications of deep learning in the video analysis of social interactions for the early identification of ASD symptoms in children aged 4 to 6 years. The study was conducted using a qualitative research synthesis method following the PRISMA guidelines. A systematic search was performed across Scopus, Web of Science, PubMed, ERIC, and specialized Persian journals for the period between 2013 and 2025. Following the application of inclusion criteria (empirical studies utilizing deep learning models to analyze social interaction videos of preschool children with or without ASD) and exclusion criteria (indirect studies, reviews without primary data), 27 eligible studies were selected. Data analysis was carried out using a two-stage coding method (open codes →\rightarrow→ concepts →\rightarrow→ core components) assisted by MAXQDA 2022 software. The quality of the included studies was appraised using the CASP checklist. From the analysis of 112 initial codes, 14 concepts and ultimately 6 key components were extracted, which are characterized by three distinguishing features: (1) focusing on objective behavioral indicators such as gaze tracking, motor coordination, and facial micro-expressions using CNN and LSTM models; (2) multimodal data fusion by combining video with neural (EEG) and physiological (GSR) signals within multi-model architectures; and (3) designing smart interactive environments based on robotics and serious games that simultaneously collect data and deliver educational interventions. Most importantly, this synthesis revealed that previous studies have predominantly focused on diagnostic accuracy (up to 94%), while seriously neglecting ethical dimensions (privacy of children’s video data), the generalizability of models to culturally diverse populations (including Iranian children), and algorithmic transparency. In conclusion, although deep learning holds transformative potential for the early diagnosis of ASD, transitioning from the laboratory to clinical practice requires the development of an ethical-pedagogical framework that integrates algorithmic objectivity with cultural sensitivities and children’s rights. Future research is recommended to: (1) develop transfer learning models to adapt Western architectures to the behavioral characteristics of Iranian children; (2) collaborate with child psychology experts to define criteria for “healthy interaction” within the Iranian population to prevent cultural bias in data labeling; and (3) design localized platforms for collecting and sharing anonymized video data of Iranian children in compliance with personal data protection regulations. Early diagnosis of autism spectrum disorder (ASD) in preschool children requires a precise analysis of social interactions, which is often prone to subjectivity and human error when using traditional methods. Deep learning, with its ability to automatically extract complex patterns from video data, offers a novel opportunity to enhance diagnostic accuracy and objectivity. This research synthesis aimed to systematically analyze the applications of deep learning in the video analysis of social interactions for the early identification of ASD symptoms in children aged 4 to 6 years. The study was conducted using a qualitative research synthesis method following the PRISMA guidelines. A systematic search was performed across Scopus, Web of Science, PubMed, ERIC, and specialized Persian journals for the period between 2013 and 2025. Following the application of inclusion criteria (empirical studies utilizing deep learning models to analyze social interaction videos of preschool children with or without ASD) and exclusion criteria (indirect studies, reviews without primary data), 27 eligible studies were selected. Data analysis was carried out using a two-stage coding method (open codes →\rightarrow→ concepts →\rightarrow→ core components) assisted by MAXQDA 2022 software. The quality of the included studies was appraised using the CASP checklist. From the analysis of 112 initial codes, 14 concepts and ultimately 6 key components were extracted, which are characterized by three distinguishing features: (1) focusing on objective behavioral indicators such as gaze tracking, motor coordination, and facial micro-expressions using CNN and LSTM models; (2) multimodal data fusion by combining video with neural (EEG) and physiological (GSR) signals within multi-model architectures; and (3) designing smart interactive environments based on robotics and serious games that simultaneously collect data and deliver educational interventions. Most importantly, this synthesis revealed that previous studies have predominantly focused on diagnostic accuracy (up to 94%), while seriously neglecting ethical dimensions (privacy of children’s video data), the generalizability of models to culturally diverse populations (including Iranian children), and algorithmic transparency. In conclusion, although deep learning holds transformative potential for the early diagnosis of ASD, transitioning from the laboratory to clinical practice requires the development of an ethical-pedagogical framework that integrates algorithmic objectivity with cultural sensitivities and children’s rights. Future research is recommended to: (1) develop transfer learning models to adapt Western architectures to the behavioral characteristics of Iranian children; (2) collaborate with child psychology experts to define criteria for “healthy interaction” within the Iranian population to prevent cultural bias in data labeling; and (3) design localized platforms for collecting and sharing anonymized video data of Iranian children in compliance with personal data protection regulations. https://www.ceciranj.ir/article_246854_0fd2aa653d00f848d82afe27000fe007.pdf