# Artificial intelligence support for diagnosis of neurodevelopmental disorders during childhood: an umbrella review

## TL;DR

This review examines how artificial intelligence can help diagnose childhood neurodevelopmental disorders, finding promising accuracy but highlighting the need for better research methods.

## Contribution

The study provides a comprehensive synthesis of AI applications for diagnosing neurodevelopmental disorders in children through an umbrella review of systematic reviews and meta-analyses.

## Key findings

- AI models achieved diagnostic accuracy ranging from 66% to 99% using data like neuroimaging and motion sensors.
- Most studies (80%) were rated as critically low in methodological quality, with only 5% achieving high quality.
- Autism spectrum disorder and ADHD were the most studied conditions, with machine and deep learning models being most commonly used.

## Abstract

The growing demand for earlier diagnosis of neurodevelopmental disorders has boosted critical assessment of artificial intelligence (AI) as a complementary tool for clinical decision-making.

This umbrella review aimed to synthesize the available evidence from systematic reviews and meta-analyses on the use of AI to diagnose during childhood any neurodevelopmental disorder [autism spectrum disorder (ASD), attention-deficit/hyperactivity disorder (ADHD), intellectual disability, communication disorders, developmental coordination disorder, and specific learning disorders]. A systematic search was conducted on the Web of Science, PsycINFO, and PubMed, covering studies published from January 2015 to August 2025 and available in any language.

Of the 148 records identified, 64 studies were included based on the predefined inclusion and exclusion criteria. ASD (n = 31) and ADHD (n = 14) were the most frequently examined conditions in which AI was applied for diagnostic purposes. To a lesser extent, it was applied to specific learning disorders (n = 5) and other developmental disorders (intellectual disability and communication disorders, jointly addressed along with other diagnoses, n = 9). The most employed AI models were machine learning (support vector machines and artificial neural networks) and particularly deep learning (such as convolutional neural networks). These models were applied to diverse data modalities, such as neuroimaging (n = 59 studies), electrophysiological (n = 19), clinical/sociodemographic (n = 15), and motion/sensor-based data (n = 11). Overall, these AI models achieved diagnostic accuracy levels ranging from 66% (based on head/facial/eye movements) to 99% (based on neuroimaging, voice, motion, and sensors). However, the methodological quality of most studies was rated as critically low according to the AMSTAR-2 criteria (80%), while only 5% of studies achieved high quality levels (focused on ASD and ADHD).

AI shows promising potential for supporting biomarker identification and diagnosis of neurodevelopmental disorders. However, future clinical implementation still requires methodologically rigorous research addressing current limitations: insufficient external validation, lack of standardization in data collection and model development, as well as reporting inconsistencies.

https://www.crd.york.ac.uk/PROSPERO/view/CRD420251110825, identifier CRD420251110825.