Health databases and early identification of autism spectrum disorder: a scoping review

Abstract

Objectives

To map the available evidence on the use of health databases for the early identification of autism spectrum disorder (ASD) across diverse populations and settings.

Design

Scoping review conducted in accordance with the Joanna Briggs Institute framework.

Data sources

Searches were conducted in MEDLINE, Embase, Scopus, PsycINFO, Web of Science and Latin American and Caribbean Health Sciences Literature, with grey literature searched through ProQuest Dissertations and Theses, up to December 2024, and updated on March 2026.

Eligibility criteria

Studies including individuals with diagnosed or suspected ASD were considered without restrictions on age, country or healthcare setting. Early identification was defined as the detection of ASD-related features prior to formal diagnosis. Health databases included digital sources such as electronic health records and surveillance systems, excluding those based solely on biological or genetic data.

Data extraction and synthesis

Study selection and data extraction were performed independently by two reviewers. Extracted data were synthesised descriptively according to database type, analytical methods and reported outcomes.

Results

Thirty-seven studies were included, mostly from high-income countries and involving children or adolescents. Data sources included electronic medical records (n=25), automated healthcare databases (n=6) and national surveys or data sets (n=6); only one database was publicly accessible. Analytical methods comprised machine learning (ML) (n=14), predictive modelling (n=6), natural language processing (NLP) (n=7) and diagnostic code algorithms (n=7). Common early predictors reported across studies include male sex, advanced maternal age, immigrant background, low SES, perinatal complications, and language or motor delays. Reported comorbidities included epilepsy, attention-deficit/hyperactivity disorder, mood/anxiety disorders and sensory issues. Children later diagnosed with ASD had more frequent medical visits, hospitalisations and emergency care. Diagnostic delays were more pronounced among low-income and underrepresented groups, despite early parental concerns.

Conclusion

Health databases have been increasingly used to support early ASD detection through methods like ML and NLP; however, their clinical application is limited by challenges related to standardisation, ethical considerations, feasibility and data access. Addressing these barriers is essential to support inclusive, evidence-based strategies.

Trial registration number

Open Science Framework. DOI: 10.17605/OSF.IO/RMVWE.

STRENGTHS AND LIMITATIONS OF THIS STUDY.

• The review followed a rigorous and transparent methodology based on the Joanna Briggs Institute framework and reported according to Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews guidelines.

• A comprehensive and librarian-validated search strategy was applied across multiple databases and grey literature sources.

• Study selection and data extraction were performed independently by two reviewers, with discrepancies resolved by consensus, ensuring methodological consistency.

• The exclusion of genetic and biomarker databases may have narrowed the scope of evidence considered.

• Most included studies originated from high-income countries, which may limit the generalisability of the findings to low-income and middle-income settings.

Introduction

Autism spectrum disorder (ASD) is a neurodevelopmental condition caused by differences in brain function and characterised by impairments in communication, social interaction and behaviour. While some cases have identifiable genetic causes, the exact origins remain unclear. Research suggests that multiple factors interact to influence neurological development, leading to atypical patterns.¹

ASD affects a significant portion of the global population. According to WHO, approximately 1 in 100 children has ASD, although prevalence estimates vary across studies and regions.² In the USA, data from the Centers for Disease Control and Prevention indicate that, in 2020, one in 36 children aged 8 years was diagnosed with ASD, representing a prevalence of 2.8% in this age group.³ In many low-income and middle-income countries, the prevalence of ASD remains largely unknown due to the absence of comprehensive epidemiological studies.²

Early identification of ASD may support earlier diagnostic processes and the development of individualised interventions.⁴ However, despite growing global awareness, the diagnosis of ASD remains a major challenge.⁵ Delayed identification can lead to more pronounced impairments, limit access to evidence-based interventions, reduce opportunities for appropriate educational placement and increase the use of psychotropic medications.^{4 6} This issue is especially evident in low-income and middle-income countries, where limited healthcare infrastructure, restricted access to diagnostic services and a shortage of trained professionals further hinder early detection and timely intervention.⁷

In addition to childhood diagnosis challenges, ASD is frequently under-recognised in adulthood. Evidence indicates that many individuals receive an ASD diagnosis several years after their first contact with mental health services, often with substantial delays between initial evaluation and diagnostic confirmation. These delays are associated with factors such as overlapping psychiatric symptoms, difficulties in reconstructing developmental history, and the presence of compensatory or camouflaging behaviours, particularly among individuals with higher cognitive abilities and among female individuals. Misdiagnosis is also common, with individuals often receiving alternative psychiatric diagnoses before ASD is correctly identified.⁸

While most research focuses on ASD in childhood, increasing attention has been given to diagnoses in adults. Autistic adults have higher rates of psychiatric disorders, such as depression, anxiety, bipolar disorder, obsessive-compulsive disorder, schizophrenia and suicide attempts.⁹ In addition, individuals with ASD frequently present with medical comorbidities, which may remain underdiagnosed due to communication difficulties, atypical symptom presentation and barriers to clinical examination, contributing to unmet healthcare needs.¹⁰ They also experience a higher incidence of medical conditions, including immune dysfunctions, gastrointestinal disorders, sleep disturbances, epilepsy and obesity.¹¹

Several real world challenges contribute to delayed or missed diagnosis of ASD. Diagnostic assessment requires detailed developmental history, which may be unavailable or unreliable, particularly in adults. Standardised diagnostic tools may have reduced sensitivity in certain populations, especially female individuals, and symptoms may overlap with other psychiatric conditions, complicating differential diagnosis. These limitations reinforce the need for complementary approaches that can support earlier and more consistent identification.⁸ Health databases have emerged as valuable tools to support early diagnosis, identify risk factors, guide healthcare strategies and predict outcomes and costs.¹² Among them, electronic health records (EHRs) stand out for their ability to provide detailed patient-level information, enabling refined clinical characterisation of individuals with ASD. When these records are standardised through structured documentation and clinical language, they can improve diagnostic reliability across settings and facilitate automated assessments in both individual-level and population-level contexts.¹³

The use of EHRs and large-scale health data for early identification has also been explored in other clinical conditions, such as psychosis risk detection and prediction of mental health crises, demonstrating the potential of these approaches to identify risk patterns and support early intervention strategies.^{14 15}

Advancements in data science, particularly in machine learning (ML) and natural language processing (NLP), have expanded the ability to extract clinically relevant patterns from large and complex health data sets, including both structured records and free-text notes.^{16 17} ML has shown promise in identifying subtle and multidimensional patterns within EHRs that may signal early signs of ASD, offering a level of analytical depth that often exceeds traditional statistical methods. These approaches support more efficient, accurate and scalable screening strategies.¹⁸

Although individual studies have examined the use of health databases for various ASD-related purposes, no synthesis has yet mapped how these databases have been used to support early identification of the condition across different populations and geographical contexts. Given the emerging and heterogeneous nature of this field, a scoping review was considered the most appropriate approach to map the extent, characteristics and methodological diversity of the available evidence.

Given the emerging and heterogeneous nature of this field, a scoping review was considered the most appropriate approach to address this question. This scoping review sought to answer the following review question: How have health databases been used to support the early identification of ASD across diverse populations and settings? The review aimed to map the available evidence and identify knowledge gaps related to the use of health databases for the early identification of ASD.

Methods and analysis

Study design

This scoping review was conducted following the methodological framework recommended by the Joanna Briggs Institute¹⁹ and was reported in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews checklist.²⁰ The objectives, inclusion criteria and methods were specified in advance and documented in a publicly available protocol, registered in the Open Science Framework (OSF). Registration number: DOI 10.17605/OSF.IO/RMVWE (https://osf.io/rmvwe/).

Eligibility criteria

Participants

Studies evaluating individuals diagnosed with or suspected of having ASD, regardless of sex, race or age group, were included.

Concept

The review considered studies that examined the use of health databases for the early diagnosis of ASD. For this purpose, a health database was defined as a data source designed for the collection, storage and analysis of large-scale information derived from electronic medical records, health surveillance systems, population databases or other digital repositories used to identify factors associated with ASD diagnosis. Databases focusing exclusively on laboratory, genetic or biomarker data sets without associated clinical or demographic health data were excluded. Early diagnosis was defined as the identification of characteristics associated with ASD prior to formal clinical confirmation.²¹

Context

No restrictions were applied regarding the geographical location of the studies or the level of healthcare setting (primary, secondary or tertiary).

Search strategy

The following databases were searched: MEDLINE (via PubMed), Embase, Scopus, PsycINFO, Web of Science and (Latin American and Caribbean Health Sciences Literature; via the Virtual Health Library portal). Grey literature was explored through ProQuest Dissertations and Theses. The search strategy combined controlled vocabulary (eg, Medical Subject Headings and Emtree) with free-text terms, as detailed in online supplemental material 1. The search strategy was validated by an experienced information specialist.

To identify additional evidence, the reference lists of all included studies and relevant scoping or systematic reviews were manually screened. The initial search covered publications up to December 2024, with no restrictions on language or publication date. To ensure that the review included the most recent evidence, the search was updated on May 2026 using the same search strategy across all databases and information sources.

Study selection and data extraction

Study selection was performed by independent pairs of reviewers. Title and abstract screening were conducted using Rayyan after the automatic removal of duplicates. Full-text screening was carried out in Microsoft Excel, with reasons for exclusion systematically documented.

Data extraction was also conducted independently by pairs of reviewers to ensure methodological rigour and consistency. A pretested data extraction form was used to maintain standardisation and accuracy in data collection. Extracted data were organised according to the Population, Concept, and Context framework, consistent with the review’s objectives and research questions. Variables included study characteristics (design, location, population, age range and sample size), database features (name, access type, purpose, data content and coverage), diagnostic and methodological aspects (criteria for ASD diagnosis, early identification methods and analytical techniques) and key findings (main results, measures of effectiveness, disparities and study-reported limitations) (see online supplemental material 2).

A third reviewer resolved any discrepancies that arose during screening or data extraction. Prior to each phase, a pilot calibration exercise was conducted, and reviewers were considered calibrated once the team achieved at least 80% agreement during title and abstract screening, full-text selection and data extraction.

Data categorisation and analysis

After data extraction, studies were categorised according to similarities in study characteristics and the types of databases used. Countries were further classified based on the World Bank income classification, which groups nations according to their average per capita income.²² Descriptive statistics were used to summarise the findings, and all analyses were performed using Microsoft Excel.

Patient and public involvement

None.

Results

The search retrieved 6614 records. After removing duplicates, 4638 records underwent title and abstract screening. Of these, 60 full-text articles were assessed for eligibility, resulting in the inclusion of 37 studies in this scoping review (figure 1). A detailed list of excluded studies, along with the reasons for exclusion, is available in online supplemental material 3.

Figure 1. Flow diagram of the study selection process.

The 37 included studies investigated the use of different types of health databases for the early identification of ASD. Most studies were conducted in high-income countries, including the USA (n=18),^{16 23–39} Canada (3),^{13 40 41} the UK (3),^42–44 Israel (6),^{17 45–49} Norway (1),⁵⁰ Sweden (1),⁵¹ Japan (1),⁵² Australia (1)⁵³ and Oman (1).⁵⁴ Only two studies were conducted in upper-middle-income and lower-middle-income countries, specifically Brazil⁵⁵ and India.⁵⁶ Online supplemental table 1 provides a comprehensive overview of the included studies.

Most studies analysed children and adolescents, although one included participants up to 24 years of age and another did not specify the age range. The types of databases varied, with electronic medical records being the most frequently used (n=25). Automated healthcare databases (n=6) and national surveys or data sets (n=6) were also employed. These databases covered regional, national and international populations with varying levels of accessibility, most requiring restricted access through preauthorised research protocols or institutional affiliations. Only one publicly accessible database was identified: the Autism Brain Imaging Data Exchange (ABIDE) in Australia, which includes neuroimaging, demographic and phenotypical data.

The methods used to identify ASD were diverse. Several studies applied ML algorithms and predictive models to analyse clinical patterns extracted from EHRs, administrative health data and research data sets, which were used to support ASD identification. NLP was also used in some studies to extract diagnostic criteria from clinical notes, supporting the identification of ASD-related features. Other studies combined conventional diagnostic classifications (eg, Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition or the International Classification of Diseases, 10th Revision) with automated data extraction techniques for early ASD identification.

The main factors associated with the early identification of ASD and the performance of the data analysis methods used in the included studies are summarised in online supplemental table 2.

Several factors were identified as relevant to the early diagnosis of ASD, including demographic characteristics, perinatal and neonatal conditions, developmental milestones, comorbidities and healthcare utilisation patterns.

Among demographic factors, male sex was consistently reported associated with ASD diagnosis. In addition, advanced maternal age (≥40 years) and immigrant background were associated with a higher risk. Socioeconomic status (SES) also influenced diagnosis: children from higher SES backgrounds were more likely to present with milder ASD phenotypes, whereas those from lower SES backgrounds were more often diagnosed with global developmental impairments.

Commonly studied perinatal and neonatal factors included prematurity (<37 weeks), caesarean delivery, low birth weight, low Apgar scores and absence of breastfeeding, all of which were associated with an increased risk of ASD. Multiple births (twins or triplets) were also identified as additional risk factors.

Developmental milestones were frequently reported as relevant indicators for early identification of ASD. Language delays were commonly associated with ASD, particularly difficulties in forming two-word sentences, limited vocabulary (<10 words) and delayed speech. Motor and cognitive deficits, such as challenges in using cutlery, engaging in play activities and performing fine motor tasks, were also associated with ASD. Parental concerns about developmental delays and hearing impairments frequently emerged as early warning signs of the condition.

With regard to comorbidities, children with ASD showed a higher prevalence of epilepsy, attention-deficit/hyperactivity disorder, mood disorders, schizophrenia and anxiety disorders. Sensory deficits, including strabismus and nystagmus, were also more common in this group. Other frequently reported conditions included sleep disturbances, feeding difficulties and regulatory problems.

The use of healthcare services was another relevant aspect. Children later diagnosed with ASD had more frequent medical visits, particularly with neurologists, psychiatrists and developmental paediatricians. They also had higher rates of hospitalisations and emergency department visits, often related to neuropsychiatric conditions. In some populations, delays in diagnostic referrals were observed, especially among low-income families and underrepresented ethnic groups, even when parental concerns were documented.

Regarding data analysis methods, approaches based on EHRs and ML models were used to identify ASD-related patterns and support data-driven identification approaches. NLP techniques were applied to extract diagnostic features from clinical notes. Studies assessing screening tools such as the Modified Checklist for Autism in Toddlers, the Early Screening of Autistic Traits, and the Young Autism and Other Developmental Disorders check-up Tool for 18-Month-Olds reported different performance measures and screening outcomes across populations and settings.

Discussion

The findings of this scoping review highlight the relevance of health databases as tools to support the early identification of ASD, particularly through the use of EHRs, automated healthcare data sets and national surveillance systems. These resources provide a structured means of tracking developmental and clinical patterns over time, which may support the identification of early signs of ASD.

The risk factors identified in this review are consistent with those reported in previous meta-analyses on ASD, including prenatal, perinatal and postnatal factors.⁵⁷

An increasing number of studies have shown that integrating individual-level data, such as demographic characteristics, perinatal factors and developmental milestones, with optimised ML algorithms have been used to support the early identification of ASD.⁵⁸ Most studies in this field originate from high-income countries, where robust technological infrastructure and standardised data collection practices support the development of data-driven approaches. This geographical concentration underscores global disparities in access to digital health resources and limits the generalisability of findings to low-income and middle-income settings.

Approaches based on large-scale EHRs and ML have been applied to analyse routinely collected clinical data and identify patterns associated with ASD diagnosis. These methods may support earlier identification of ASD-related features, contributing to clinical decision-making as a complementary approach to clinical assessment.¹⁴

Although these findings highlight the potential of health databases and ML to support early ASD identification, they should be interpreted with caution, as none of the included studies directly compared the diagnostic performance with standardised diagnostic instruments such as the Autism Diagnostic Observation Schedule or the Autism Diagnostic Interview–Revised. Consequently, this review does not allow conclusions regarding their comparative diagnostic performance. Instead, the available evidence suggests that these approaches are best viewed as complementary to existing diagnostic pathways, supporting earlier recognition of individuals who may benefit from comprehensive clinical assessment.

An important consideration in predictive modelling is the variability in how predictors are measured across different types of data systems, ranging from granular clinical records to broader administrative databases. Inconsistencies in data quality, definitions and population characteristics may affect model performance and limit generalisability to other contexts.⁵⁹ In addition to these methodological considerations, issues related to equity and model performance were identified.

While this review provides a valuable synthesis of studies using health data for early ASD detection, it also highlights a critical gap related to algorithmic bias and the generalisability of predictive models. Many of the included studies did not stratify analyses by sociodemographic subgroups or evaluate potential disparities in model performance. This omission is particularly concerning, as predictive algorithms may unintentionally perpetuate or exacerbate existing healthcare inequities, especially when trained on data that reflect structural biases and systemic disadvantages.⁶⁰

Another major limitation identified in this review is the lack of external validation. Few studies evaluated whether their models performed consistently across different healthcare environments or population groups. This is particularly important for applications in low-resource settings, where data quality, healthcare delivery and population characteristics may differ substantially from the original context of model development.⁶¹ Furthermore, none of the included studies referenced the TRIPOD (Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis) statement, a widely recognised framework developed to enhance the clarity, reproducibility and clinical relevance of predictive modelling research.⁶² Adopting TRIPOD standards in future research would improve methodological transparency and ensure that predictive tools are rigorously developed, well-documented and suitable for real world implementation.

Despite their potential, most databases used in the included studies were not publicly accessible. Access was typically restricted to researchers affiliated with specific institutions or subject to controlled research protocols. This limitation creates barriers to replication, external validation of predictive models and wider dissemination of findings. Notably, the only publicly available database identified, ABIDE,⁶³ includes neuroimaging data alongside demographic and phenotypical information rather than exclusively clinical or administrative data. Expanding access to de-identified, ethically governed health data sets could foster more inclusive and generalisable research across diverse settings.

Finally, although the application of ML to health databases has been explored for advancing the early identification of ASD, its implementation in real world healthcare remains challenging.^{59 60} Ethical concerns, data privacy regulations, legal frameworks and equity considerations continue to pose significant barriers. In practice, the data driving these models are often heterogeneous, incomplete and poorly standardised, which may affect model accuracy and reproducibility.⁶⁴ Many ML models also lack transparency and interpretability, making it difficult for clinicians to understand or trust their outputs, particularly when external validation has not been performed. These challenges highlight the need for transparent, ethically grounded and equity-oriented approaches when applying artificial intelligence tools in healthcare.^{59 65}

This review has some methodological limitations. As a scoping review, it did not include critical appraisal of the methodological quality of the included studies, which limits the assessment of the strength of the evidence. Most included studies were conducted in high-income countries, which may restrict the generalisability of the findings to other contexts. In addition, the heterogeneity of data sources, definitions and analytical methods may affect the comparability of the results. The limited availability and restricted access to most databases reported in the included studies may hinder replication and external validation of predictive models. Furthermore, limitations related to external validation and the assessment of algorithmic bias were identified across the included studies.

Conclusion

Recognising autism at an early stage is essential for ensuring timely support. The findings of this review demonstrate that health databases, particularly those incorporating ML and NLP, have been increasingly explored to support early ASD identification across diverse populations and settings. However, their translation into clinical practice remains limited by challenges related to standardisation, ethical considerations, technological feasibility and data accessibility. Expanding access to high-quality, de-identified data and promoting transparent, equity-oriented modelling practices will be important for advancing inclusive and evidence-based strategies for early ASD identification. Future studies should prioritise the external validation of predictive models across different populations and healthcare systems to strengthen reproducibility and real world applicability.

Data availability statement

Data are available upon reasonable request. All data relevant to the study are included in the article or uploaded as supplementary information.