Artificial intelligence for autism spectrum disorder: advances in diagnosis, behavior analysis and educational support

The findings of this systematic review highlight the growing expansion of artificial intelligence (AI) applications in the field of Autism Spectrum Disorder (ASD), reflecting a progressive transition from exploratory technological experimentation toward increasingly structured approaches with potential clinical and educational applicability. Overall, the studies analyzed suggest that artificial intelligence is being integrated into four major domains: early detection and diagnostic support, automated behavioral and social analysis, educational intervention, and communication support. However, rather than representing a homogeneous field of development, the literature reveals substantial methodological diversity, technological variability, and differing levels of empirical validation. Consequently, the interpretation of the available evidence requires a critical and integrative perspective capable of examining not only the technological promise of AI systems but also their methodological robustness, practical applicability, limitations, and ethical implications (Table 4).

One of the principal findings emerging from the present review concerns the heterogeneity of artificial intelligence approaches applied to ASD. The reviewed literature includes classical machine learning models, deep learning architectures, computer vision systems, natural language processing tools, and multimodal approaches integrating several sources of behavioral data. Importantly, these technological approaches differ considerably in their methodological requirements, interpretability, predictive performance, and potential applicability within clinical and educational contexts.

Studies based on classical machine learning approaches, including those reported by Abbas et al. (2019), Thabtah (2019), Thabtah and Peebles (2020), and Alharbi et al. (2022), primarily relied on structured behavioral datasets to classify characteristics associated with ASD. These models often used behavioral questionnaires or predefined variables to identify patterns related to autism symptomatology. One advantage of traditional machine learning approaches lies in their relative interpretability, allowing researchers and professionals to better understand the variables contributing to prediction outcomes. In clinical contexts, interpretability represents an important advantage because healthcare professionals require transparency in diagnostic support systems, particularly when decisions affect access to specialized intervention or educational accommodations. However, despite their promising predictive capabilities, many of these studies relied on relatively small datasets and limited external validation procedures, which raises concerns regarding generalizability across diverse populations and contexts.

In contrast, deep learning approaches demonstrated increasing relevance, particularly in studies involving video analysis and behavioral pattern recognition. Kim et al. (2025), for example, illustrate how deep learning systems can analyze home videos to identify behavioral indicators associated with autism with high levels of predictive performance. Similarly, computer vision–based approaches examined by Washington et al. (2020, 2021) focused on the automated extraction of behavioral markers related to social interaction, visual attention, and eye contact. Compared with traditional machine learning, deep learning systems may achieve superior predictive accuracy due to their capacity to process large volumes of complex and unstructured data, including audiovisual information. Nevertheless, these advantages are accompanied by important methodological and practical challenges. Deep learning models often require extensive training datasets and substantial computational resources, while simultaneously presenting lower levels of interpretability. In clinical and educational environments, the “black box” nature of some AI systems may limit trust among practitioners and families, particularly when decision-making processes remain insufficiently transparent.

The increasing incorporation of explainable artificial intelligence frameworks represents an important development in this regard. Jiang et al. (2025) demonstrate how explainable AI approaches may improve transparency by clarifying how predictive decisions are generated through multimodal behavioral information. From a practical perspective, explainability may be especially important in ASD contexts because diagnostic and educational decisions frequently involve multidisciplinary teams requiring interpretable and accountable evidence. Consequently, explainable AI may represent an important pathway toward the responsible integration of machine learning systems into professional practice.

Natural language processing (NLP) approaches also represent a relevant emerging line of research in the field of autism. Bone et al. (2020) show that linguistic markers can be analyzed through AI systems to identify communication patterns associated with ASD. Similarly, Wang et al. (2023) report advances in AI-assisted speech generation systems capable of supporting communication processes in children with autism. Compared with visual or behavioral analysis systems, NLP-based technologies may offer unique opportunities to improve expressive communication and facilitate social interaction, particularly among individuals experiencing language-related difficulties. However, linguistic variability across developmental stages, communication profiles, and cultural contexts represents an important challenge for the generalizability of these systems.

Another relevant observation emerging from this review concerns the methodological robustness and validation procedures of AI models applied to ASD. Although several studies reported promising levels of predictive performance, including high accuracy rates and favorable classification indicators, substantial variability was identified regarding training procedures, validation methods, sample composition, and outcome measures. Importantly, predictive performance should not be interpreted in isolation from methodological rigor. Several studies relied on relatively small or highly specific samples, which may increase the risk of overfitting and reduce the external validity of findings. Models trained on highly homogeneous datasets may perform effectively under controlled research conditions while demonstrating limited applicability in real-world clinical or educational environments.

Moreover, validation procedures varied considerably across studies. While some investigations incorporated independent datasets or cross-validation methods, others relied primarily on internal testing procedures. This variability complicates direct comparison among studies and limits the capacity to determine which approaches demonstrate superior effectiveness under real-world conditions. The lack of methodological standardization across datasets, behavioral indicators, and evaluation metrics further complicates synthesis of findings and reinforces the need for common frameworks capable of improving comparability across future investigations.

In this regard, the methodological quality assessment conducted in the present review revealed that, although most studies demonstrated moderate to high quality, several recurrent limitations persisted. Among the most common concerns identified were restricted sample sizes, limited external validation, dataset imbalance, heterogeneity of performance indicators, and insufficient longitudinal evidence. These findings are particularly relevant because the successful implementation of AI systems in ASD contexts requires not only predictive effectiveness but also stability, replicability, and reliability across different populations and settings.

Beyond methodological robustness, an equally important consideration concerns the practical applicability of artificial intelligence systems in real clinical and educational environments. Across the studies analyzed, artificial intelligence was predominantly conceptualized as a complementary support mechanism rather than a substitute for professional judgment. This distinction is especially relevant in the field of Autism Spectrum Disorder, where diagnosis typically relies on multidisciplinary evaluation processes incorporating developmental history, behavioral observation, standardized instruments, and clinical expertise. Accordingly, the reviewed studies suggest that artificial intelligence systems may serve primarily as screening, classification, and decision-support tools capable of complementing existing professional practices.

Within the domain of early detection and diagnostic support, several studies indicate that machine learning algorithms may facilitate the identification of behavioral indicators associated with autism, particularly in contexts where access to specialized assessment remains limited. Tariq et al. (2019, 2022), for example, demonstrate how digital tools based on machine learning may analyze home videos to identify early behavioral patterns associated with ASD. Similarly, Abbas et al. (2019), Thabtah and Peebles (2020), Rasul et al. (2024), and Jiang et al. (2025) emphasize the growing capacity of predictive models to support diagnostic screening and behavioral classification processes. Nevertheless, although these findings demonstrate considerable technological potential, important caution remains necessary regarding interpretation and implementation. Most of the reviewed studies validated AI systems against already established clinical instruments, including gold-standard procedures such as the Autism Diagnostic Observation Schedule, Second Edition (ADOS-2), and the Autism Diagnostic Interview-Revised (ADI-R). Consequently, artificial intelligence systems should not be interpreted as autonomous diagnostic tools but rather as complementary mechanisms capable of supporting earlier identification, reducing waiting times, and assisting professional decision-making processes.

This distinction becomes especially important when considering the clinical implications of classification errors. False positive classifications may generate unnecessary concern among families and increase pressure on already constrained diagnostic services, potentially leading to inappropriate referrals and unnecessary psychological stress. Conversely, false negative outcomes may delay access to early intervention services during critical developmental periods, thereby limiting opportunities for timely educational and therapeutic support. In this regard, the implementation of AI-based systems in ASD assessment requires careful balancing between predictive efficiency and clinical responsibility. Rather than replacing clinical expertise, artificial intelligence appears most beneficial when functioning within hybrid models that integrate algorithmic prediction with multidisciplinary professional evaluation.

The reviewed literature also suggests substantial educational potential for artificial intelligence systems designed to support learning and social participation among students with ASD. Chen et al. (2021), Hu and Han (2022), and Lan et al. (2024) report promising results regarding adaptive educational environments capable of personalizing learning experiences according to students’ cognitive and behavioral characteristics. Intelligent tutoring systems and AI-supported educational technologies may facilitate individualized instruction, adapt task complexity, provide immediate feedback, and promote social participation within inclusive educational contexts. These characteristics are particularly relevant for students with autism because educational trajectories frequently require differentiated pedagogical responses capable of addressing highly heterogeneous profiles of communication, attention, sensory regulation, and social interaction.

However, despite this promising potential, the reviewed studies also reveal important barriers limiting educational implementation. Several interventions were evaluated within highly controlled research settings, often involving relatively small samples and short intervention periods. Consequently, evidence regarding long-term effectiveness and sustainability in real-world classrooms remains limited. Educational implementation additionally requires adequate technological infrastructure, teacher training, interdisciplinary collaboration, and institutional support, all of which vary considerably across educational systems and socioeconomic contexts. Therefore, although AI-based educational systems may contribute to more personalized learning environments, further validation remains necessary to establish their practical effectiveness within ordinary school settings.

Another important issue emerging from this review concerns equity, bias, and the generalizability of artificial intelligence systems applied to autism. Several authors emphasize that machine learning models are inherently dependent on the quality and representativeness of the datasets used during training procedures. Consequently, biases embedded within datasets may become reproduced or amplified through algorithmic systems. This issue appears particularly relevant in Autism Spectrum Disorder research because ASD populations frequently present substantial demographic, developmental, and cultural heterogeneity.

Most reviewed studies relied on datasets characterized by limited diversity in terms of age, ethnicity, socioeconomic background, and clinical presentation. Additionally, autism research has historically shown a predominance of male participants, despite increasing recognition that girls and women may present distinct behavioral and communicative manifestations often associated with delayed or missed diagnosis. Consequently, AI systems trained primarily on male-dominated datasets may risk underperforming when applied to female populations or individuals presenting less conventional symptom profiles. Similarly, the overrepresentation of clinical or urban populations may reduce applicability in rural or underserved contexts where behavioral characteristics, healthcare access, and environmental conditions may differ substantially.

The challenge of generalizability therefore represents one of the principal limitations of current AI-based autism research. Although predictive models frequently demonstrate promising levels of performance under experimental conditions, their effectiveness may decrease substantially when transferred to heterogeneous real-world environments. Future research should therefore prioritize the development of larger, more diverse, and culturally representative datasets capable of supporting more equitable and generalizable AI systems. Particular attention should also be given to developmental variability across childhood, adolescence, and adulthood, given that behavioral manifestations of autism evolve considerably over time.

Ethical and regulatory concerns also emerge as central issues regarding the integration of artificial intelligence into autism-related contexts. The increasing use of behavioral datasets, home videos, wearable sensors, facial recognition systems, and communication data inevitably raises questions related to privacy, consent, transparency, and accountability. Several reviewed studies relied on audiovisual materials collected within domestic environments, which may involve substantial concerns regarding confidentiality and data security. Families participating in these technologies may not always possess a full understanding of how behavioral information is processed, stored, or reused by algorithmic systems.

Furthermore, concerns surrounding algorithmic transparency remain highly relevant. The limited interpretability of some machine learning and deep learning systems complicates understanding of how predictive outcomes are generated, potentially undermining trust among clinicians, educators, and families. In response to these concerns, explainable artificial intelligence approaches such as those proposed by Jiang et al. (2025) may represent an important direction for future development by improving transparency and facilitating professional interpretation of AI-generated recommendations.

Finally, several limitations of the present review should be acknowledged. First, the heterogeneity of study designs, technological approaches, outcome measures, and validation procedures limited direct comparison among studies and prevented quantitative synthesis through meta-analysis. Second, although a structured methodological quality assessment was conducted using Joanna Briggs Institute critical appraisal tools, variability in methodological quality across included studies may influence interpretation of findings. Third, this systematic review was not prospectively registered in PROSPERO or another international registry. Although the review protocol was developed following PRISMA 2020 methodological recommendations, the absence of prospective registration may limit methodological transparency and should therefore be considered when interpreting the findings.

When positioning the present review within the broader scientific literature, an important distinction should be emphasized. Recent systematic reviews and meta-analyses have primarily focused on specific AI-related domains within Autism Spectrum Disorder, particularly diagnostic prediction models, behavioral monitoring systems, and machine learning approaches based on home-video analysis. In contrast, the present review adopted a broader interdisciplinary perspective by simultaneously examining four major domains of AI application in ASD: early detection and diagnostic support, behavioral and social analysis, educational intervention, and communication support technologies. This broader scope provides a more integrated understanding of how artificial intelligence is progressively being incorporated across clinical, educational, and social dimensions of autism support, thereby extending beyond narrower diagnostic or behavioral frameworks commonly represented in recent reviews.

An important contribution of the present review lies in its broader interdisciplinary scope. Unlike previous reviews primarily centered on diagnostic prediction models or behavioral monitoring systems, the present study integrates evidence across four complementary domains of AI application in Autism Spectrum Disorder: early detection and diagnostic support, behavioral and social analysis, educational intervention, and communication support technologies. This integrative perspective is particularly relevant because individuals with ASD frequently require coordinated responses across clinical, educational, and social environments. Consequently, understanding artificial intelligence exclusively from a diagnostic perspective may offer only a partial understanding of its broader potential. By incorporating educational and communication-related applications alongside diagnostic and behavioral approaches, the present review provides a more comprehensive framework for understanding how AI may contribute to more personalized, adaptive, and inclusive forms of autism support.

Overall, the evidence synthesized in this review suggests that artificial intelligence represents a rapidly expanding and highly promising field for advancing the understanding, detection, educational support, and communication opportunities of individuals with Autism Spectrum Disorder. Nevertheless, the reviewed literature simultaneously demonstrates that technological innovation alone is insufficient to guarantee meaningful implementation. Future progress in this field will depend on the development of methodologically robust, transparent, ethically responsible, and clinically interpretable systems capable of responding to the diversity and complexity of autistic populations. Greater emphasis on external validation, equitable dataset construction, interdisciplinary collaboration, and long-term educational and clinical evaluation will be essential to ensure that artificial intelligence contributes not only to technological advancement but also to genuinely inclusive and evidence-based support for individuals with autism.