Applying artificial intelligence in neurodevelopmental disorders management and research

Abstract

Artificial intelligence (AI) is increasingly being used in the diagnosis, treatment, and monitoring of neurodevelopmental disorders, enabling earlier detection, personalised interventions, and continuous support. Traditional machine-learning models such as logistic regression, random forests, and support vector machines remain valuable for their interpretability and their ability to integrate multimodal clinical data. Deep-learning (DL) approaches, including convolutional neural networks and transformer-based architectures, improve the analysis of neuroimaging and behavioural datasets and strengthen diagnostic and prognostic performance. Important challenges remain, including limited transparency in DL systems, ongoing concerns about data privacy and algorithmic bias, and a lack of large and diverse paediatric datasets that restricts generalisability. Interpretability tools such as SHAP and LIME offer partial solutions but still lack standardised evaluation. At the same time, AI-driven robotic platforms are enhancing therapeutic engagement and supporting skill acquisition in children with neurodevelopmental conditions. This review highlights that AI tools have strong potential to act as clinical adjuncts rather than replacements, providing earlier detection, personalised management, and scalable care models. Realising this potential will require rigorous validation, ethical safeguards, and thoughtful integration into human-led care pathways.

Introduction

Neurodevelopmental disorders are a group of conditions that arise during early development and are characterised by difficulties in cognition, communication, motor skills, and behaviour. Common diagnoses within this group include autism spectrum disorder (ASD), attention-deficit hyperactivity disorder (ADHD), intellectual disability, and developmental language disorder. These conditions frequently present with overlapping symptoms and may co-occur, complicating both diagnosis and intervention [1]. In addition, symptom expression varies widely between individuals and across contexts, further increasing diagnostic complexity. Although behavioural observations and standardised assessments are routinely used in clinical practice, they may not fully capture the subtle or context-specific challenges experienced by children [1]. Delays in diagnosis, particularly in conditions, such as ADHD, are well documented and can limit access to early intervention and support services [2]. Conversely, early identification and intervention in disorders such as ADHD and ASD have been associated with improved social, emotional, academic, and functional outcomes [2–5].

Given the complexity of neurodevelopmental disorders, there is increasing interest in new approaches that can support both research and clinical practice. One such approach is artificial intelligence (AI), which refers to computer-based methods that analyse data, identify patterns, and assist with prediction or decision-making. In medical research and care, AI has been used to improve diagnostic accuracy, support outcome prediction, and facilitate the analysis of large patient datasets [6]. These applications are particularly relevant in neurodevelopmental disorders, where overlapping symptoms continue to complicate diagnosis and management [7, 8]. Despite its potential, the application of AI in neurodevelopmental disorders also raises important limitations and risks. A key concern is algorithmic bias. Evidence suggests that AI language models may systematically associate neurodivergent terms, such as “autism” or “ADHD,” with negative concepts including disease or danger [9]. This is particularly concerning in a field already affected by disparities in diagnosis and access to services across different populations.

This review summarises the fundamental principles of AI and its broader applications in the early diagnosis, personalised intervention and monitoring of neurodevelopmental conditions. It also critically assesses the associated technical, ethical, and systemic challenges. Finally, it discusses future directions, emphasising the importance of transparency, accessibility and responsible innovation in the development of AI-driven clinical care for neurodevelopmental conditions.

Methodology

This narrative review was conducted using the Scale for the Assessment of Narrative Reviews (SANRA) [10]. The SANRA framework informed the structure and reporting of the review, including the literature search strategy, selection of relevant studies, narrative synthesis of findings, and critical discussion of methodological limitations.

Data sources and search strategy

The literature search was conducted using databases such as PubMed/Medline, Scopus, Embase, IEEE Xplore, CINAHL and the Cochrane Library. Specific keywords were used, including ‘artificial intelligence’, ‘machine learning’, ‘deep learning’, ‘neurodevelopmental disorders’, ‘autism spectrum disorders’, ‘attention deficit hyperactivity disorder’, ‘motor disorders’, ‘learning disabilities’ and ‘communication/speech disorders’. Also Medical Subject Headings (MeSH) and Boolean operators (AND and OR) were used to structure the queries in the databases, as demonstrated by the following: (("artificial intelligence"[Mesh] OR "artificial intelligence"[tiab] OR "machine learning"[Mesh] OR "machine learning"[tiab] OR "deep learning"[tiab] OR "neural network*"[tiab] OR "natural language processing"[tiab] OR "computer vision"[tiab])) AND (("Neurodevelopmental Disorders"[Mesh] OR "Neurodevelopmental disorder*"[tiab] OR "Autism Spectrum Disorder"[Mesh] OR "autism spectrum disorder*"[tiab] OR ASD[tiab] OR "Attention Deficit Disorder with Hyperactivity"[Mesh] OR ADHD[tiab] OR "communication disorder*"[tiab] OR "speech disorder*"[tiab] OR "language disorder*"[tiab] OR "learning disability*"[tiab] OR dyslexia[tiab] OR "motor disorder*"[tiab] OR dyspraxia[tiab])) AND (child*[tiab] OR pediatric*[tiab] OR paediatric*[tiab] OR adolescent*[tiab]). This approach ensured that the literature search targeted our specific area of interest. In addition, a manual search was conducted to identify references to recently published case-specific reviews, providing further insight into the field of digital health and neurodevelopmental disorders.

Article selection and assessment

The inclusion criteria permitted various study designs, including observational studies, case–control studies, cohort studies, randomised controlled trials, systematic reviews, and meta-analyses. Studies focusing on paediatric populations were prioritised. The review included articles published in English only, with the timeframe for selecting these papers being from inception until 2025. Conference abstracts without full papers, non-peer reviewed studies and preprints and blog posts were excluded from the review.

The records were screened at the title, abstract and full text levels by a six-member team (M.F., S.M., A.B.J., S.R., S.I. and A.A.M.). Most articles were eliminated from consideration at the title and abstract levels. Any disagreements during the screening process were resolved through discussion, after which the senior author (A.A.W.) made the final decision. Of the initial 9,390 articles provided by the databases following application of the search strategy, 3,080 were scrutinised after the exclusion of 6310 duplicates. Ultimately, 159 studies were eligible for inclusion in the main results analysis of this study following title, abstract, and full-text screening and subsequent exclusions for various reasons (see Fig. 1). The results were synthesised using narrative analysis.

Fig. 1.

A flow chart for the article selection process. The literature search was conducted using the databases; PubMed/Medline, Embase, Scopus, IEEE Xplore and CINAHL. Of the 9390 initial articles provided by the databases following application of the search strategy, 3080 were scrutinised after 6310 duplicates were excluded. Ultimately, 168 studies were eligible for inclusion in the main results analysis of this study after title and abstract level (n = 1159) and full-text level (n = 1753) exclusions

An overview of AI and its application in paediatric neurology

In the context of healthcare, AI encompasses diverse approaches including machine learning (ML), artificial neural networks (ANNs) and deep learning (DL). To further understand how AI can be applied in paediatric neurology, it is important to have a grasp on what ML, DL and ANN are within the AI umbrella. AI is the broadest concept, it refers to machines or systems that can perform tasks that would normally require human intelligence, like problem-solving, ML can be described as a subset of AI. It is a method that allows machines to learn from data without being explicitly programmed. Instead of being told what to do, ML algorithms find patterns in data and use them to make predictions or decisions [11, 12]. DL is a subset of ML that uses ANNs designed to mimic the way the human brain processes information. These networks are made up of many layers hence the word "deep" which allows them to learn complex patterns from large amounts of data [13]. Figure 2 summarises and explains the relationship between AI, ML and DL.

Fig. 2.

Illustration of the hierarchical structure of AI, ML, and deep learning DL. AI encompasses all computational methods designed to perform tasks that typically require human intelligence. ML forms a subset of AI focused on learning patterns directly from data to generate predictions or decisions. DL is a further subset of ML that uses multilayered neural networks capable of automatically extracting complex features from large, high-dimensional datasets. Together, these layers show how increasingly specialised techniques build on one another within modern AI. (Created with BioRender.com)

In the field of paediatric neurology and neuroscience, AI holds a lot of promise due to the unique challenges in this field. This is because children often struggle to communicate symptoms clearly, and neurodevelopmental disorders frequently present with overlapping and evolving clinical features that can make diagnosis difficult at times [8]. In paediatric neurological disorders, predictive models are increasingly utilising a range of data types such as neuroimaging, clinical information, and genetic profiles in order to improve the predictive accuracy.

As with most health conditions, early diagnosis is critical in neurodevelopmental disorders, as timely interventions can significantly improve developmental trajectories and long-term outcomes [14]. ML and DL, when trained on radiomics platforms such as structural magnetic resonance imaging (MRI) features, can effectively predict, detect and classify neurodevelopmental disorders and their subtypes in children, enabling early diagnosis and treatment [14–16]. Early studies have examined how neonatal MRI could predict neurodevelopmental outcomes in very preterm infants at two years of age [17]. They found that structural abnormalities, particularly in white matter, were strongly associated with subsequent cognitive and motor impairments. While ML was not directly employed in the original study, the integration of neuroimaging and clinical data yielded essential features that are now routinely incorporated into ML models. Subsequent research has expanded upon this foundation by using algorithms to automatically analyse comparable datasets; thereby, enhancing the prediction of risks and the planning of interventions for preterm infants. More recently, ML methods have become adept at integrating diverse datasets, such as imaging clinical and genetic data, to improve prediction accuracy [18].

In paediatric seizure and epilepsy care, ML and DL techniques using electroencephalogram (EEG)- and non-EEG-based methods are employed to improve the diagnosis and monitoring of epileptic seizures [19, 20]. In order to address the need for an automated seizure classification system, deep neural network (DNN) models such as seizures, periodic and rhythmic continuum patterns deep neural networks (SPaRCNet) have been designed to classify ictal–interictal–injury continuum events. These models have been shown to perform better than most experts based on calibration and discrimination metrics [21]. In addition to classification, ML and DL are also used to monitor responses to anti-seizure medications [22]. Furthermore, DL models have shown promise in enabling wearable, real-time systems for predicting seizures [23].

AI has also recently been used to predict migraine and non-migraine headaches in children and adolescents. In one study, an AI-based model using 17 objective criteria was developed to distinguish between migraine and non-migraine headaches in children aged 6–17 [24]. The model was trained and internally validated using a dataset comprising 909 patients who had been diagnosed according to the International Classification of Headache Disorders (ICHD-3). The model demonstrated high diagnostic accuracy for paediatric and adolescent migraine [24].

Although AI has made significant progress in many areas of paediatric care, its application in paediatric neuro-oncology is still in its infancy. Nevertheless, general AI models are increasingly being used to diagnose brain tumours, particularly for tumour classification and detection [25–28]. ML algorithms such as support vector machines (SVMs), logistic regression, k-nearest neighbours (KNNs), random forests (RFs), extreme gradient boosting (XGB) and neural networks, combined with MRI-based radiomics, have been used to classify paediatric brain tumours [28]. Furthermore, DL architectures such as the residual neural network (ResNet), ResNeXt and DenseNet have demonstrated high accuracy in classifying posterior fossa tumour pathologies using T2-weighted MRIs [27]. Overall, AI technologies are progressively enhancing diagnostic accuracy and helping to overcome the limitations of conventional methods in the diagnosis of paediatric brain tumours [29–31].

When used appropriately, AI has the potential to transform paediatric neurology by enabling more accurate, timely and personalised care. In everyday clinical settings, it can help neurologists by making diagnostic testing more efficient, supporting personalised treatment planning and encouraging better adherence to therapy. These systems can also be used to develop, anticipate, and reduce treatment-related side effects, which could greatly improve the overall quality of life for children and their families.

The application of AI in the management of neurodevelopmental disorders

Autism spectrum disorder

ASD is a neurodevelopmental condition that typically emerges within the first three years of life. It is referred to as a “spectrum” due to the wide variability in type and severity. According to the Diagnostic and Statistical Manual of Mental Disorders, 5th edition (DSM-5) [32], individuals with ASD may exhibit restricted, repetitive interests and behaviours, along with difficulties in social interaction and academic functioning. Early detection and intervention are critical to addressing the neurodevelopmental needs of affected individuals and achieving optimal outcomes [33].

Globally, an estimated 61.8 million people, around one in 127 individuals, were on the autism spectrum in 2021, with prevalence highest in high-income Asia Pacific regions such as Japan (≈ 1587 per 100,000) and lowest in countries like Bangladesh (≈ 588 per 100,000) and China (≈ 656 per 100,000). Rates in the Middle East and North Africa were intermediate (≈ 772 per 100,000), underscoring marked regional variation linked to diagnostic capacity and awareness. ASD ranks among the top ten causes of non-fatal health burden in individuals under 20 years, reinforcing the global need for early screening and sustained support services [34].

In the current era of expanding AI applications in healthcare, the use of AI and its subfields, ML and DL, in the prediction, prevention, and management of ASD has gained significant attention [35, 36].

Prediction and early diagnostics of ASD

AI enables earlier ASD detection, with diagnosis before age two linked to improved neurodevelopmental outcomes such as higher IQ scores [37]. Although ASD has no cure, early identification allows timely behavioural and environmental interventions [38]. AI tools now use behavioural, imaging, and physiological data to support screening and diagnosis [39–41].

Neuroimaging-driven DL models, such as the Xception framework, have demonstrated the ability to detect subtle cortical and subcortical alterations associated with ASD [42]. At the same time, comparative studies indicate that classical machine learning (ML) models, including logistic regression, SVM, and KNN, can perform competitively or outperform DL approaches in behavioural datasets [43]. These findings highlight the value of multimodal and algorithm–agnostic strategies for early ASD detection.

Furthermore, an AutMedAI (XGBoost) model achieved an area under the receiver operating characteristic curve (AUC-ROC) of 0.895, a sensitivity of 0.805, a specificity of 0.829 and a positive predictive value (PPV) of 0.897 in a cohort of 30,660 participants, identifying developmental milestones and eating behaviour as the top predictors [44]. Another multicentre study combining The autism diagnostic interview-revised (ADI-R) and the autism diagnostic observation schedule (ADOS) data from 162 children outperformed traditional screening tools such as the modified check- list for autism in toddlers (M-CHAT) and child behavior checklist (CBCL) [40]. ML classifiers, including logistic regression, SVC, KNN, and random forest, achieved 100% accuracy in children aged 4 to 11. On toddler data, the logistic regression model achieved an intersection over union (mIoU) of approximately 99.8%, indicating near-perfect classification. When federated learning was applied to datasets comprising more than 600 cases per site, the accuracy rate remained high at 98% [45]. The mIoU indicates how well the model’s predictions align with the true diagnostic labels. Although this metric is commonly used in image analysis to compare predicted and actual regions, in this study it was applied to demonstrate the accuracy with which the model classified toddlers by diagnostic status. An mIoU of around 99.8% shows that the model's predictions were almost identical to the confirmed diagnoses, reflecting excellent performance.

DL further improved diagnostic precision. The ResNet152 achieved 89% accuracy, while a ResNet-Vision Transformer hybrid reached 91.33% accuracy, 91.89% recall, and 90.79% specificity [46]. Behavioural DNN ensembles achieved 88% sensitivity and 94% specificity [47]. A review of 39 EEG and magnetoencephalography (MEG) studies confirmed that SVM and DL models were most effective for ASD detection and symptom classification [48].

Complementing these findings, it has been demonstrated that sophisticated DL architectures trained on neuroimaging datasets can capture high-dimensional ASD biomarkers, while ML techniques remain highly effective even when relying solely on behavioural feature sets [42, 43]. Together, these studies reinforce the growing consensus that early ASD prediction benefits from model diversity, multimodal data integration, and cross-validated evaluation across large cohorts.

The recent models demonstrate ongoing progress. The S/SD-ASD model achieved an accuracy of 97.5% using masked autoencoders and attention-based self-distillation [49]. The Multi-Head convolutional neural network (CNN) with BERT (MCBERT) model, which integrates bidirectional encoder representation from transformers (BERT) and CNNs, achieved 93.4% accuracy, and the bio-inspired WS-BiTM model reached 97.6% accuracy [50]. Additional optimisation-driven approaches further support the early detection of ASD. The AFF-BPL model uses an adaptive feature fusion pipeline that combines Bat and Particle Swarm Optimisation with an LSTM classifier. This achieves 99.2% accuracy in toddlers, 98.9% in children and 98.6% in adults while reducing overfitting and improving generalisability across datasets [51]. Together, these findings demonstrate the increasing precision and generalisability of AI in the prediction of early ASD.

Early intervention and personalised treatment of ASD

AI applications in early intervention for children with ASD include intelligent educational systems for personalised learning, real-time monitoring platforms, and support tools that enhance adaptive functioning [39]. A 12-month observational study of 43 children aged 2–18 years evaluated the CognitiveBotics platform alongside standard therapy [52]. Participants were assessed at baseline and follow-up using the Childhood Autism Rating Scale (CARS), Vineland Social Maturity Scale, Developmental Screening Test, and Receptive-Expressive Emergent Language Test. In the intervention group, consistent AI platform use produced reductions in CARS scores (mean 33.41 to 28.34) and gains in social age, developmental age, and receptive and expressive language. The control group, with inconsistent use, showed no significant changes. Structured caregiver training included web-based sessions and home guidance [52].

AI-assisted technologies also address sensory and behavioural challenges in classrooms and homes. One system integrated sensor fusion and ML to identify environmental triggers of sensory distress and used fuzzy logic to generate real-time alerts and personalised regulation strategies, improving engagement and usability [53]. A broader review of AI-driven technologies, including robots, virtual reality (VR), and mobile platforms, found improvements in social skills (47%), daily living (26%), and communication abilities (16%) among children with neurodevelopmental disorders [53]. In a randomised clinical trial of 103 infants showing early ASD signs, pre-emptive behavioural intervention reduced the rate of ASD diagnosis by age three (7% in the intervention group versus 21% with usual care) [54].

AI serving as intelligent monitoring systems for dynamic tracking and meeting the diverse needs of patients with ASD

AI-powered intelligent monitoring systems dynamically track and support individuals with ASD, enhancing daily functioning, personalised care, and safety management. Systems combining wearable wristbands with physiological sensors, emotion-recognition cameras, and RFID-enabled interactive learning tools collect comprehensive data processed by AI models to monitor emotional and physical states. These models generate adaptive audiovisual feedback and provide real-time updates to caregivers through connected mobile applications [55].

ML algorithms have been applied to large-scale developmental and behavioural datasets to improve ASD monitoring and identification. The Autism and Developmental Disabilities Monitoring Network used a random forest classifier trained on developmental evaluation texts, achieving 86.5% concordance with clinician diagnoses [56]. Systems integrating environmental and physiological data employed Gradient Boosting Decision Tree and Random Forest models to predict attention and stress levels in children with ASD, reaching accuracies of 86.67% and 99.05%, respectively [54]. These approaches enable scalable monitoring and timely intervention across diverse settings. Recent multimodal architectures also enhance continuous monitoring by integrating behavioural, physiological, and imaging-derived meta-features. The MCBERT framework combines a multi-head CNN with a BERT encoder to fuse MRI data and clinical characteristics, achieving 93.4% accuracy in leave-one-site-out testing and demonstrating strong stability across heterogeneous acquisition environments [50].

Real-time AI monitoring also enhances safety. The Autistic Child Sensor and Assistant system uses wireless sensor networks and ML to detect children’s gestures and movements, issuing immediate alerts to caregivers [57]. Similarly, the AutiLife system combines fifth generation connectivity with SVM algorithms to monitor physiological signals such as heart rate and body temperature, enabling rapid emergency responses in autism centres [58]. Figure 3 illustrates the key strengths of the application of AI in ASD.

Fig. 3.

Overview of how AI supports the management of ASD. AI assists early detection by analysing behavioural cues, developmental data, imaging, and physiological signals. It also contributes to personalised intervention through intelligent educational systems, real-time monitoring platforms, and tools that help address sensory, communication, and behavioural challenges. In daily life, AI-enabled monitoring systems provide adaptive feedback, support functional skills, and deliver timely updates to caregivers across home and classroom settings. (Created with BioRender.com). ASD Autism spectrum disorder, AI artificial intelligence, DL deep learning

Critical appraisal: strengths, limitations, and research gaps of AI applications in ASD

AI tools for ASD show strong potential but continue to face important design and real-world challenges. The AutMedAI model achieved high accuracy in identifying ASD but depended on retrospective parent-reported data, which introduces recall bias and limits control over key factors. Its drop in accuracy on external testing also suggests that it was over-fitted to its original dataset and included too little demographic diversity [44]. More recent models that combine behavioural and biological information show similar issues, with performance differences between groups, indicating that variations in genetics and demographics still limit how well these systems can be generalised [59].

Physiological and behavioural pipelines also face reproducibility limits. Eye-tracking classifiers predict ASD risk accurately in small, age-restricted cohorts but lack ecological validation [60]. EEG-based ML frameworks report high accuracy yet rely on under-powered datasets with inconsistent preprocessing [61]. DL analyses of an autism brain imaging data exchange functional MRI data similarly vary with scanner and preprocessing differences, showing that harmonised acquisition and external replication remain prerequisites for clinical credibility [62].

New text-based AI models have widened the possibilities for autism screening but often repeat language-related biases found in their training data. Hybrid systems that combine natural language processing (NLP) and ML achieve strong results within their own datasets but show inconsistent performance across different languages and cultures, underscoring the need for fairness auditing [63]. While early behavioural therapy reduces ASD diagnosis rates by age 3, no study yet confirms that AI-driven triage accelerates access or improves long-term outcomes [64]. For now, AI should remain a clinician-supervised triage adjunct, pending multicentre fairness-validated trials linking model predictions to measurable developmental gains [59, 61, 64]. Even models reporting very high accuracy face generalisation constraints when applied across populations with varying behavioural profiles. For example, although the WS-BiTM framework achieved 97–98% accuracy through White–Shark-based feature selection and Bi-LSTM classification, its own authors emphasised challenges related to dataset size, heterogeneity, and the risk of overfitting despite optimisation steps [65]. Table 1 summarises the models and datasets used, as well as the key outcomes of AI in ASD.

Table 1.

Overview of AI models used in ASD research, including their application areas, datasets, and reported performance

AI model/system	Application domain	Dataset/participants	Performance metrics (as reported)	Key findings/outcomes	Main limitations (as stated)
Parent + Clinician ML Classifiers (ADI-R + ADOS) [40]	Combined screening tool integration	162 children (multicentre)	Improved accuracy vs M-CHAT and CBCL (traditional tools)	Combining parent and clinician inputs enhanced screening accuracy	Small sample; external replication needed
Xception-based DL Model [42]	Neuroimaging-based ASD diagnosis	Neuroimaging dataset (MRI scans; sample size not explicitly stated in abstract)	High diagnostic classification performance using Xception; model captured fine-grained cortical/subcortical features (exact accuracy not provided in open abstract)	Demonstrated that Xception architectures effectively extract ASD-related neurofunctional patterns; supports early detection through imaging biomarkers	Limited reporting of dataset size; generalisability constrained by single-modality input (neuroimaging only); requires larger, diverse cohorts for validation
Comparative ML Models; SVM, Random Forest, Neural Networks, Logistic Regression [43]	Behavioural-feature-based ASD classification	Multiple ASD behavioural datasets; empirical comparison across models (precise sample sizes dependent on dataset, not detailed in abstract)	Multiple ML models showed high accuracy, robustness, and stability across datasets; comparative metrics indicated strong performance across classifiers	ML classifiers demonstrated consistent performance across datasets; shows effectiveness of diverse algorithmic approaches for ASD screening	Limited details on dataset size and demographic distribution; lack of external validation; performance dependent on feature quality rather than multimodal integration
AutMedAI (XGBoost) [44]	Prediction using clinical and developmental features	30 660 participants; 28 variables (eating habits, sleep, milestones)	AUC-ROC 0.895; Sensitivity 0.805; Specificity 0.829; PPV 0.897	Identified developmental milestones and eating behaviour as top predictors	Retrospective parent-reported data; recall bias; limited demographic diversity; over-fitting on external testing
ML models (Logistic Regression, SVC, Decision Tree, KNN, Random Forest) [45]	Prediction of ASD presence/absence in children and toddlers	Children aged 4–11 and toddler datasets	100% accuracy (children and toddlers); mIoU ≈ 99.8 (log reg)	High diagnostic accuracy across age groups	Evaluated on limited datasets; requires independent replication
AFF-BPL Model (Bat + PSO + LSTM) [51]	Cross-age ASD prediction using adaptive feature-fusion optimisation	Toddlers, children, and adults (multi-dataset evaluation)	Accuracy: 99.2% (toddlers), 98.9% (children), 98.6% (adults)	Adaptive feature-fusion reduced overfitting, improved cross-dataset generalisability	Requires high computational cost; external clinical validation not yet done
WS-BiTM Framework (White-Shark Optimisation + Bi-LSTM) [65]	ASD prediction using behavioural and developmental features with optimised sequencing	Behavioural/developmental ASD datasets (moderate size)	Accuracy 97–98%	White-Shark feature selection enhanced Bi-LSTM learning efficiency	risks of overfitting, small sample sizes, and reduced generalisability
Federated Learning (SVC + Logistic Regression) [45]	Cross-site ASD prediction	Four datasets > 600 cases each	Accuracy 98%	Shows privacy-preserving multi-site training feasibility	No reported clinical deployment
DL Models (ResNet152, DenseNet201, VGG16, EfficientNet-B0, MobileNetV2, VGG19, ViT hybrid) [46]	Image-based diagnosis in children	Image datasets of ASD children	ResNet152 89%; ResNet152 + ViT 91.33% accuracy; Precision 90.67%; Recall 91.89%; Specificity 90.79%	Transfer-learning and hybrid models improved classification accuracy	Dataset bias; limited demographic diversity and ecological validation
MCBERT Multimodal Framework (CNN + BERT) [50]	MRI + clinical feature fusion for ASD prediction; multimodal continuous monitoring	Multi-site MRI datasets with clinical/behavioural metadata	Accuracy 93.4% (leave-one-site-out cross-validation)	Stable performance across heterogeneous scanners and acquisition sites	High computational demand; MRI-dependent framework limits universal scalability
Behavioral Biomarker Identification Framework (DNN + Ensemble) [47]	Behavioural pattern classification	ASD behavioural datasets	Sensitivity 88%; Specificity 94%	Effective behaviour-based ASD classification	Details of cohort size not reported
EEG/MEG ML Classifiers (SVM, DL Architectures) [48]	EEG and MEG signal analysis	Systematic review of 39 studies	Not quantified (broadly high accuracy reported)	Used for ASD detection, severity prediction, and cognitive state classification	Non-standardised methods; inconsistent pre-processing across studies
CognitiveBotics Platform [52]	Early intervention/adaptive learning	43 children (2–18 years); 12 months	↓ CARS 33.41 → 28.34; ↑ language and social scores	Improved language and social maturity vs control group	Small sample; non-randomised design
Sensor-Fusion System (fuzzy logic) [53]	Sensory distress detection in classroom	Pilot evaluation	Improved engagement and usability	Generated real-time alerts and personalised strategies	Prototype; short-term assessment
AI Therapeutic Robots/VR Tools [53]	Adaptive social and communication training	Systematic review of AI technologies	Social skills ↑ 47%; daily living ↑ 26%; communication ↑ 16%	Enhanced adaptive functioning via AI tools	Short trials; heterogeneous outcome metrics
Pre-emptive Behavioural Intervention (RCT) [54]	Early behavioural therapy for infants at risk of ASD	103 infants (12-month RCT)	ASD diagnosis 7% vs 21% (control)	Reduced ASD diagnosis rate by age 3	AI role not isolated from therapist intervention
GBDT/Random Forest Models [54]	Stress and attention prediction using physiological data	Children with ASD	Accuracy 86.67–99.05%	Predicted attention and stress levels accurately	Closed algorithms; small samples
Wearable + Camera + RFID Platform [55]	Emotion and activity monitoring	Prototype deployment	Not quantified – real-time monitoring described	Provided adaptive feedback and alerts to caregivers	Hardware cost; privacy concerns
ADDM Monitoring Network (Random Forest) [56]	Automated diagnosis via developmental records	National registry data	86.5% concordance with clinician diagnoses	Efficient text-based classification pipeline	Dependent on record quality
Autistic Child Sensor and Assistant [57]	Gesture/movement safety alerts	Prototype system	Immediate parent alerts (recorded qualitatively)	Improved safety monitoring via sensors	Reliance on hardware accuracy
AutiLife (5G + SVM) [58]	Physiological emergency monitoring in autism centres	Prototype implementation	Not quantified – continuous monitoring reported	Rapid response enabled by 5G connectivity	No large-scale clinical validation

The table summarises how different ML and DL approaches have been applied for diagnosis, behavioural assessment, early intervention, and real-time monitoring. Key findings highlight improvements in prediction accuracy, identification of clinically relevant behavioural patterns, and support for personalised intervention. Reported limitations mainly relate to small samples, dataset bias, inconsistent preprocessing, and limited clinical validation

AUC-ROC area under the receiver operating characteristic curve, PPV Positive Predictive Value, RFID Radio-Frequency Identification, mIoU Mean Intersection-Over-Union, CARS Childhood Autism Rating Scale, RCT randomised controlled trial, ADDM Autism and developmental disabilities monitoring network, AI artificial intelligence, DL deep learning, SVM Support Vector Machine, ViT vision transformer, DNN deep neural network, ML machine learning, VR virtual reality, ADDM Autism and developmental disabilities monitoring, RFID radio frequency identification, GBDT Gradient-Boosted Decision Trees, CNN convolutional neural network, ASD Autism spectrum disorder, EEG electroencephalogram, MEG magnetoencephalography, MRI magnetic resonance imaging, 5G fifth generation, M-CHAT modified checklist for autism in toddlers, CBCL child behavior checklist, PPV positive predictive value, XGBoost extreme gradient boosting, ResNet residual neural network

Attention-deficit/hyperactivity disorder

Enhancing ADHD screening and diagnosis

AI has significantly advanced the early and precise identification of ADHD, a neurodevelopmental disorder often characterised by hyperactivity, inattention and impulsivity. Traditional diagnosis relies on clinical interviews and standardised scales, which can be limited by observer bias, diagnostic delays and cultural variability [66]. By contrast, AI frameworks, including supervised learning models (e.g. random forests and SVMs), CNNs and decision-tree ensembles, can process high-dimensional neuroimaging, behavioural and genetic data more objectively and with greater predictive power [7, 67, 68].

Several systematic reviews emphasise the superior performance of AI in multimodal data integration. The recent eXplainable-AI (XAI) frameworks have reported similarly high diagnostic performance in both children and adults, with logistic regression and XGB models achieving up to 99–100% accuracy in population ADHD datasets [69]. For instance, ML applications were examined in psychometric data and found that diagnostic accuracy exceeded 90%, surpassing traditional scales such as the Conners Rating Scale[66]. Furthermore, EEG and MEG biomarkers processed via AI pipelines (e.g. XGBoost) and recurrent neural networks (RNNs) have facilitated subtype classification (e.g. inattentive versus combined ADHD) and trait-based symptom clustering [70, 71]. Complementary EEG-decomposition frameworks using STFT and LightGBM have recently reached 91–96% accuracy for paediatric ADHD detection, highlighting rapid progress in electrophysiological AI biomarkers [72]

AI-enhanced VR is an emerging technology that provides immersive and ecologically valid diagnostic tools. Studies have demonstrated that integrating VR with real-time gaze tracking and reinforcement learning algorithms improves the differentiation of attentional lapses as compared to paper-based tests [73]. These findings were corroborated by broader scoping reviews, which highlight AI's enhanced sensitivity in capturing real-world attentional behaviours [74, 75]. Notably, these systems also improve diagnostic equity in low-resource settings by minimising clinician dependence [76, 77].

Aiding in personalised treatment approaches

AI facilitates the transition from one-size-fits-all paradigms to personalised, precision-based ADHD management. Using patient-specific data from electronic health records (EHRs), wearable devices and behavioural logs, AI models, especially reinforcement learning, unsupervised clustering and transformer-based models, can predict treatment responses and adapt intervention protocols accordingly [67, 78].

Several systematic reviews have reported that AI-supported pharmacological decision systems outperform traditional trial-and-error strategies. For example, AI-based meta-analyses of methylphenidate efficacy, guided by clustering algorithms, have demonstrated higher remission rates and fewer side-effects than psychiatrist-led dosage decisions [68, 74]. Such systems typically use predictive models, such as gradient boosting machines or probabilistic graphical models, to recommend the most suitable treatment options. In addition, AI-based clinical decision-support systems such as TDApp now generate personalised, evidence-based pharmacological recommendations using symbolic AI, demonstrating strong concordance with major clinical guidelines [79]. In one clustering study of 33 children undergoing methylphenidate therapy, distinct response profiles were identified, supporting a personalised medicine approach [80]. Another imaging-based deep-clustering study of 150 individuals produced biotypes that corresponded with differential medication responses [81].

Non-pharmacological interventions also benefit from AI integration. Game-based cognitive training and behavioural reinforcement platforms are now powered by AI agents that adjust the level of difficulty and rewards in real time based on user input [82, 83]. In addition, AI-powered digital therapy platforms using MEG normalisation and DL significantly reduced impulsivity in ADHD patients more effectively than therapist-guided cognitive behavioural therapy (CBT) [71].

Furthermore, large language models (LLMs) including ChatGPT-4o and robotic assistants have been incorporated into therapy routines to provide personalised psychoeducation and coaching, thereby improving treatment adherence and socio-emotional outcomes. The same study demonstrated a 93% success rate in tailored interactions and stable performance in 91% of extreme input scenarios [84].

Empowering individuals with ADHD to better self-manage their condition

AI-powered mobile health (mHealth) platforms and wearable systems are reshaping ADHD self-management by applying ML models such as anomaly detection, NLP, and behavioural clustering to enable real-time monitoring, feedback, and personalised goal setting [77, 82]. These technologies can detect behavioural patterns and environmental triggers linked to symptom fluctuations, allowing users and clinicians to adjust routines proactively [74, 76]. In one clinical study, an AI-driven digital therapeutic used with 30 children with ADHD led to a 36.84% reduction in inattention and a 50.71% decrease in hyperactivity–impulsivity after four weeks of treatment [85].

Unlike static or rule-based digital tools, adaptive AI interventions tend to keep users more engaged and consistent in their use [75]. By gradually adjusting task difficulty through reinforcement learning, these systems help sustain motivation and reduce disengagement. Some also incorporate sentiment-analysis features that monitor changes in emotional state over time [83]. A randomised controlled trial involving 104 adults is currently underway to test the Sincrolab Adults or the control group through the AI-based cognitive-stimulation programme, developed to enhance attention and executive functioning in adults with ADHD. The study reflects growing interest in personalising cognitive therapies through adaptive AI frameworks [86].

Several TensorFlow-based DL frameworks are now integrated into smartwatches and smart speakers to deliver hands-free ADHD coaching [84]. Compared with non-AI technologies such as generic behavioural apps or augmented-reality platforms, AI systems offer superior adaptability, predictive intelligence, and scalability [87, 88]. Figure 4 summarises the key strengths of the application of AI in ADHD.

Fig. 4.

Overview of how AI supports different aspects of ADHD care. AI enhances early screening by analysing complex behavioural, neuroimaging and physiological data with higher accuracy than traditional tools. It also assists personalised treatment by predicting medication response and adapting interventions using electronic health records, wearables and behavioural logs. In daily life, AI-driven mobile health systems, language models and wearable devices help individuals track symptoms, identify triggers and receive real-time coaching. (Created with BioRender.com). ADHD Attention-Deficit-Hyperactivity Disorder, AI artificial intelligence, SVM Support Vector Machine, CNN convolutional neural network, NLP natural language processing, DL deep learning

Critical appraisal: strengths, limitations, and research gaps of AI applications in ADHD

Recent AI models for ADHD show promising results but remain limited by small, single-centre datasets and validation confined to research settings. This inflates reported accuracy and restricts generalisability across scanners, age groups, and clinical environments [89, 90]. Multimodal VR tools combining eye-tracking, EEG, movement, and behavioural data provide more realistic assessments, yet most studies occur under controlled conditions rather than in routine practice [91]. Subgroup calibration for sex, medication status, and comorbidity is seldom reported, leaving fairness and stability untested [89]. These concerns are consistent with recent adult ADHD XAI studies showing that even transparent SHAP-based models require multi-site validation to ensure fairness across demographic and clinical subgroups [92].

Physiological and behavioural sensors offer objective symptom measurement but often rely on small, convenience samples and proprietary data-processing pipelines that limit reproducibility and transparency [93–95]. Although wearable systems can track medication response and day-to-day symptom change, long-term adherence and device compatibility remain poorly understood, reducing readiness for large-scale clinical use [94, 95].

Early clinical trials suggest that AI-based cognitive training may improve symptoms and brain function, but most are short, small, and lack outcome measures for real-world benefit such as reduced wait times, improved academic performance, or decreased medication use [71]. Future research should prioritise large, multi-centre prospective studies with pre-registered protocols, harmonised data formats, subgroup and fairness analyses, and outcomes that measure whether AI-guided screening or personalised care enhance daily functioning and treatment access [89–91, 93, 94]. Table 2 summarises the datasets, models, and key outcomes of AI in ADHD.

Table 2.

Summary of key AI models, datasets and reported outcomes in ADHD

AI model/system	Application domain	Dataset/participants	Performance metrics (as reported)	Key findings/outcomes	Main limitations (as stated)
Supervised ML (Random Forest, SVM, Decision-Tree Ensembles) [7, 66, 68]	Screening and diagnosis of ADHD using multimodal data	Behavioural, neuroimaging and genetic datasets	Not specified ('greater predictive power')	Objectively processed high-dimensional data; reduced observer bias	Diagnostic performance varies by data type; bias not eliminated
Explainable-AI ML Models (Logistic Regression + XGBoost) [69]	High-accuracy ADHD diagnosis in children and adults using interpretable ML	Population-level ADHD datasets (children + adult cohorts)	Accuracy 99–100%	Provided transparent feature attribution; maintained high diagnostic performance across age groups	Generalisation constrained by dataset homogeneity; requires external multi-site validation
EEG/MEG RNN and XGBoost models [70, 71]	Subtype classification and trait-based clustering	EEG and MEG datasets	Not quantified ('facilitated classification and clustering')	Distinguished inattentive versus combined subtypes	Small samples; need for standardisation
STFT–LightGBM EEG Classification Framework [72]	Paediatric ADHD detection using time–frequency EEG decomposition	Paediatric EEG datasets (multi-session recordings)	Accuracy 91–96%	strong electrophysiological discrimination for ADHD biomarkers	Limited to paediatric populations; requires validation across diverse EEG acquisition environments
VR + gaze-tracking + reinforcement learning [73–75]	Ecologically valid diagnostic testing	Experimental VR studies and scoping reviews	Not quantified ('improved differentiation of attentional lapses')	Provided better attention assessment than paper-based tests	Conducted only in research settings; limited generalisability
AI screening in low-resource settings [76, 77]	Accessible diagnosis and health equity	Field implementations	Not quantified ('improved diagnostic equity')	Reduced clinician dependence; improved accessibility	Requires digital infrastructure
AI-supported pharmacological decision systems [68, 74, 78, 100]	Medication response and dosage prediction	Meta-analyses of methylphenidate efficacy	Not quantified ('higher remission rates and fewer side effects')	Improved dose selection compared with psychiatrist-led decisions	Heterogeneous data; potential meta-bias
TDApp Symbolic-AI Decision-Support System [79]	Personalised pharmacological recommendations for ADHD	Clinical cases compared to guideline-based standards	Not quantified (“strong concordance with major clinical guidelines”)	Enabled personalised dose and drug-selection recommendations using symbolic-AI reasoning; enhanced transparency and clinical interpretability	Dependent on guideline completeness; limited quantitative validation in large diverse cohorts
AI-based cognitive training and reinforcement platforms [71, 82, 83,]	Non-pharmacological therapy	ADHD participants in training trials	Not quantified ('reduced impulsivity more than CBT')	Enabled adaptive difficulty and reward personalisation	Short duration; small sample sizes
LLM and robotic assistants (e.g. ChatGPT-4o) [84]	Therapeutic coaching and psychoeducation	Clinical and home-use trials	Qualitative ('outperformed traditional telehealth bots')	Enhanced adherence and empathy in therapy sessions	Early-stage; ethical and privacy concerns
mHealth/wearable/NLP platforms [74, 75, 77, 82, 83]	Self-management and symptom tracking	Mobile and wearable device users	Not quantified ('better adherence and engagement than rule-based apps')	Provided adaptive feedback and goal-setting	Long-term adherence and compatibility remain unclear
TensorFlow DL smartwatch/speaker assistants [84]	Hands-free coaching for children	Comparative study versus caregiver intervention	Not quantified ('superior task persistence and emotional regulation')	Delivered continuous personalised support	Small sample size; short-term evaluation
Transformer-based personalised models [87, 88]	Adaptive behavioural intervention and scalability	Conceptual and pilot frameworks	Not quantified ('demonstrated therapeutic parity and scalability')	Showed adaptability and predictive intelligence	Lacked quantitative clinical validation

The table outlines how various supervised, DL, reinforcement learning and transformer-based systems have been applied across diagnostic, therapeutic and self-management domains. It highlights the range of data sources used, including behavioural measures, neuroimaging, EEG/MEG signals, virtual reality environments, electronic health records, and mobile/wearable sensor streams, and summarises the main performance trends and reported clinical benefits. Limitations primarily relate to small samples, heterogeneous datasets, lack of standardisation, limited real-world testing and early-stage development

ADHD attention-deficit hyperactivity disorder, AI artificial intelligence, ML machine learning, SVM Support Vector Machine, EEG electroencephalography, MEG magnetoencephalography, RNN recurrent neural network, XGBoost extreme gradient boosting, VR virtual reality, LLM large language model, NLP natural language processing, DL deep learning, CBT cognitive behavioural therapy, mHealth mobile health

Learning disabilities

Identification and diagnosis of learning disabilities

Conventional diagnostic methods for learning disabilities, such as dyslexia, dysgraphia, dyscalculia, auditory processing disorder (APD) and language processing disorder (LPD) often rely on extensive clinical observation, standardised testing and teacher referrals. While these approaches have contributed greatly to the effective management of learning disabilities, they remain limited by their subjective nature, referral delays and variability in expertise across educational and clinical settings. In contrast, AI and ML offer scalable, data-driven diagnostic models that can process large and heterogeneous datasets to detect patterns, enabling earlier and more accurate identification.

In dyslexia, the researchers have explored various techniques to identify key cognitive and language markers, using tools like NLP, eye-tracking, and voice analysis. For instance, a recent scoping review showed that DL models, including CNNs and RNNs, can accurately predict dyslexia-related traits by analysing phonological and orthographic information, demonstrating strong potential for early detection in school-age children [96]. Another review highlighted AI-based dyslexia detection systems and found that models integrating multimodal data outperform traditional screening methods in terms of both accuracy and processing speed. These models often utilise minimal input (e.g. reading tasks or writing samples) to produce results that would otherwise require a series of psychometric assessments [97]. These studies suggest that such technologies could shorten the time to diagnosis while maintaining or even improving reliability.

In addition, for dyscalculia, an AI-based diagnostic tool that uses logical inference engines to identify deficits in numerical processing has been introduced [98]. This early work has been built upon to conduct a systematic evaluation of AI-enhanced dyscalculia screening methods [99]. These findings suggest that AI models trained on mathematical error patterns and cognitive task performance can achieve a level of diagnostic specificity that rivals that of human experts. Similarly, the dimensional overlap between dyslexia and dyscalculia has also been emphasised, advocating ML approaches that can model multiple cognitive traits simultaneously, a capacity that is not feasible with standard assessments [100]. In the case of dysgraphia, AI applications have centred on handwriting analysis, with supervised learning models capable of assessing character formation, spacing and writing pressure using stylus-tracking data. These systems provide non-invasive, real-time assessments that are sensitive to subtle motor-coordination patterns often missed in traditional evaluations [101].

Similarly, AI applications have advanced the early detection of auditory processing disorder and language processing disorder. One of the earliest models for detecting central auditory processing anomalies using EEG signals was developed in 2004 [102]. More recent work has compared multiple ML classifiers applied to auditory brainstem response data in children with auditory processing disorder, demonstrating superior classification accuracy over clinician-based evaluations, particularly in borderline or comorbid cases [102, 103]. Building on this, NLP has been used to identify children with language disorders through the analysis of spontaneous speech and narrative language samples, enabling objective detection of linguistic anomalies without the need for extensive testing [105].

Compared with traditional approaches, these AI-based systems can drastically reduce the time needed for diagnosis, make large-scale screening more feasible, and offer consistent decision-making. However, the generalisability of such systems is often limited by data heterogeneity and the underrepresentation of diverse populations in training datasets.

Stress detection in individuals with learning disabilities

Students with learning disabilities often experience higher stress levels due to academic pressure, social exclusion and unmet educational needs. Although traditional stress monitoring has long been reliable for stress detection and management in patients with learning disabilities, it relies on caregiver reports or observational checklists, which are retrospective and subject to bias. In contrast, AI-driven stress detection combines physiological sensing with predictive modelling to deliver real-time, objective insights.

In a recent study, a learning model trained on physiological data measures, including electrodermal activity, heart rate variability and interbeat intervals, successfully detected stress responses in individuals exposed to controlled virtual reality environments [106]. The study employed a Leave-One-Out Cross-Validation technique, demonstrating that personalised AI models significantly outperformed generalised models and random baselines across all groups, including individuals with intellectual and learning disabilities.

These findings are consistent with a study using personalised multitask learning to predict mood changes based on wearable sensor data [107]. Their model emphasised the importance of individualised baselines, highlighting the limitations of one-size-fits-all approaches. The advantage of AI lies not only in its sensitivity, but also in its ability to provide proactive alerts, enabling early intervention to prevent behavioural escalations, a functionality absent in traditional systems.

Enhancing academic skills through AI-based interventions

Conventional pedagogical strategies for learners with disabilities usually include differentiated instruction, the use of assistive technologies and regular one-to-one support. Although these approaches are effective, they are resource-intensive and often lack real-time adaptability. AI-enhanced educational technologies offer a solution by providing dynamic, personalised learning pathways based on continuous performance monitoring.

In a study conducted in Saudi Arabia, ten AI-supported intervention sessions were implemented to enhance numeracy and reading skills among Arabic-speaking students with intellectual disabilities [108]. The experimental group exhibited statistically significant improvements across all four domains of the Woodcock-Johnson IV Achievement Test. Reading scores increased from 80.5 to 88.2 and numeracy scores from 78.7 to 86.4. These gains were sustained at follow-up, and educators reported high satisfaction with the interventions, particularly emphasising the adaptability and engagement features of AI-based instruction [108]. Although the study demonstrates the promise of AI tools in special education for students with mild intellectual disabilities, further research is needed to explore long-term impacts and wider applicability across different educational settings [108].

A systematic review has shown that AI-driven interventions for dyslexia, dyscalculia, and dysgraphia achieve moderate to large effect sizes in skill acquisition. These interventions outperformed traditional remediation techniques that relied solely on repetitive drills or static software [109].

AI-powered writing support systems have also benefited learners with dysgraphia. These tools provide real-time spelling corrections and syntactic feedback, as well as multimodal prompts that are tailored to each learner's cognitive profile[101]. By contrast, conventional approaches such as occupational therapy or paper-based practice require a longer timeframe to produce similar improvements and are less scalable.

For language processing and auditory disorders, intelligent tutoring systems that use speech recognition and NLP, offer immediate feedback on pronunciation, grammar, and sentence structure. They can spot unusual patterns in language development sooner than traditional evaluations conducted by speech-language therapists [105]. These advances in computer-automated methods have created valuable opportunities to make language sample analysis more efficient and accessible, enabling faster diagnosis and more timely treatment of language disorders in children.

All this demonstrates that AI-enabled approaches are transforming the landscape of learning disability diagnosis, stress management, and academic intervention. Across disorders such as dyslexia, dyscalculia, dysgraphia, APD, and LPD, these technologies show superior scalability, personalisation and efficiency compared to traditional methods. However, future work must address model generalisability, ethical deployment, and integration with human-led educational ecosystems. Figure 5 illustrates the key strengths of the application of AI in learning disabilities.

Fig. 5.

Overview of how AI supports the identification and management of learning disabilities. AI tools assist diagnosis by analysing reading patterns, handwriting samples, language use, and numerical processing to detect dyslexia, dysgraphia, and dyscalculia. Physiological and behavioural models can detect stress responses that affect learning, enabling early intervention. AI-driven instructional tools further enhance academic skills by providing real-time feedback, adaptive prompts, and personalised support for reading, writing, and comprehension. (Created with BioRender.com). AI artificial intelligence, NLP natural language processing

Critical appraisal: strengths, limitations, and research gaps of AI applications in learning disabilities

AI has demonstrated considerable potential in the identification and understanding of learning disabilities, including dyslexia, dyscalculia, dysgraphia, APD and LPD. Across these domains, AI models have achieved diagnostic accuracies comparable to or exceeding those of traditional clinician-led methods, particularly through the integration of multimodal inputs such as linguistic, visual, and neurophysiological data [96, 97]. The capacity of DL and ML systems to uncover subtle cognitive and behavioural markers, often imperceptible to human observation, positions them as valuable tools for early detection and personalised educational planning [105].

However, despite promising results, several methodological and translational limitations constrain the robustness and applicability of current AI research. In dyslexia, while CNN- and RNN-based models have shown strong predictive capacity for phonological and orthographic deficits [96, 97], most datasets comprise small, homogeneous, and monolingual samples. This narrow linguistic and cultural representation restricts the generalisability of findings across diverse orthographic systems. Moreover, the opacity of DL architectures limits interpretability, which in turn diminishes clinician trust and complicates the integration of AI outputs into established psychometric assessment frameworks. Direct comparative studies between AI predictions and clinician diagnoses remain rare, leaving questions about diagnostic equivalence unresolved.

In dyscalculia for instance, AI-based inference engines and error-pattern recognition models have advanced significantly [98, 99]. These systems demonstrate sensitivity to numerical and spatial reasoning anomalies comparable to expert evaluation. Nonetheless, data scarcity, heterogeneous sampling across developmental stages, and the lack of longitudinal validation hinder their translation into practice [67]. Few studies have examined whether early AI-assisted detection of dyscalculia leads to measurable improvements in long-term learning outcomes or educational attainment. In addition, most work remains confined to experimental contexts rather than real-world classrooms, limiting ecological validity.

For dysgraphia, supervised ML models trained on handwriting features such as letter formation, spacing, and writing pressure show clear advantages in scalability and non-invasive screening [101]. Yet, current datasets are largely collected through stylus-based digital platforms, which may not accurately reflect handwriting behaviour on traditional paper. The limited representation of diverse age groups, languages, and writing systems poses further barriers to generalisability. Furthermore, the majority of models rely solely on kinematic data, neglecting multimodal signals such as linguistic content or fine-motor coordination, which could enhance diagnostic specificity.

In the case of APDs and LPDs, AI models leveraging EEG, auditory brainstem responses, and NLP techniques have achieved impressive classification accuracy [102, 105]. However, most studies remain limited to controlled laboratory settings, thereby constraining ecological validity and real-world generalisation. The underrepresentation of children from diverse linguistic, cultural, and socioeconomic backgrounds introduces significant bias, which may affect diagnostic reliability across populations. Ethical challenges surrounding the acquisition, storage, and interpretation of sensitive neural and speech data are also insufficiently addressed in the literature. Furthermore, while prototype models demonstrate technical efficacy, their integration into clinical workflows and educational systems remains nascent, highlighting a persistent gap between computational innovation and practical deployment.

Despite notable progress, several critical research gaps remain across AI applications for learning disabilities. A primary limitation lies in the restricted dataset diversity and representativeness, as most studies employ small, monolingual, and culturally homogeneous samples, reducing generalisability across orthographies, age groups, and educational systems [96, 101]. The lack of longitudinal research further constrains understanding of developmental trajectories and the sustained impact of early AI-based detection on learning outcomes [99, 100]. In addition, most current models are evaluated in highly controlled laboratory or experimental environments, limiting ecological validity and practical translation into real-world classroom or clinical settings [102, 105].

Another persistent challenge involves the opacity of DL architectures, which impedes interpretability and clinician trust, thus restricting integration with established diagnostic frameworks [96, 97]. Moreover, few studies have systematically compared AI predictions with traditional psychometric or clinician-led assessments, leaving diagnostic equivalence largely unverified [98, 100]. Finally, ethical and data governance concerns, particularly those related to neural and speech data privacy, remain insufficiently addressed [102, 105]. Addressing these research gaps will require large-scale, multilingual datasets, explainable and transparent AI models, and sustained interdisciplinary collaboration among computer scientists, linguists, clinicians, and educators to ensure both technical robustness and equitable implementation. Table 3 summarises the studies that discuss the AI models and datasets used, and the key outcomes in relation to learning disabilities.

Table 3.

Summary of AI models, datasets, and reported outcomes across learning disabilities

AI Model/System	Application Domain	Dataset/Participants	Performance Metrics (as reported)	Key Findings/Outcomes	Main Limitations (as stated)
CNNs & RNNs [96]	Dyslexia detection via phonological and orthographic feature analysis	School-age children; primarily English-speaking monolingual datasets	Accuracy > 90% in predicting dyslexia-related traits	Demonstrated strong predictive capability for early dyslexia detection; effective use of multimodal inputs (text, audio, eye-tracking)	Small, homogeneous samples; limited cross-linguistic generalisability; lack of interpretability in deep models
Multimodal DL system [97]	Dyslexia screening using linguistic, reading, and writing data	Children screened through reading/writing samples	Outperformed traditional screening in speed and reliability	Reduced diagnostic time; improved screening efficiency	Underrepresentation of diverse populations; absence of clinician-AI comparison studies
Logical inference engine [98]	Dyscalculia detection via logical and numerical reasoning	Early developmental cohort; controlled task-based data	Diagnostic precision comparable to clinicians	Detected subtle reasoning deficits through symbolic inference	Limited to experimental data; small samples
AI-enhanced screening [99]	Dyscalculia screening through ML-based mathematical error analysis	Students with suspected dyscalculia	F1-scores > 0.85; accuracy > 88%	Identified numerical and spatial reasoning deficits with high precision	Lack of longitudinal validation; limited cross-age generalisation
ML model for multidimensional traits [100]	Overlapping dyslexia–dyscalculia modelling	Mixed learning disability dataset	Diagnostic accuracy 86%	Modelled multiple cognitive traits simultaneously; captured comorbid features	Small-scale validation; limited real-world classroom testing
Supervised ML handwriting model [101]	Dysgraphia detection via handwriting features (spacing, pressure, formation)	Students using stylus-based handwriting tasks	Classification accuracy ~ 89%	Enabled real-time, non-invasive dysgraphia assessment	Stylus-only data limits ecological validity; lacks multimodal integration
EEG-based neural network [102]	APD detection	Children with central auditory deficits	Early neural signature detection accuracy 80–85%	Established proof-of-concept for EEG-based CAPD detection	Controlled laboratory setting; limited clinical translation
ML classifiers for ABR signals [103, 104]	APD diagnosis using ABR data	Paediatric cohorts with APD and controls	Classification accuracy > 90%	Outperformed clinician-based evaluations; improved borderline case detection	Homogeneous datasets; limited ecological validity; ethical issues in data collection
NLP-based language anomaly detection [105]	LPD detection from spontaneous speech	Children with developmental language disorders	Sensitivity ~ 88%; specificity ~ 85%	Automated identification of linguistic anomalies; enabled rapid, objective screening	Underrepresentation of multilingual speakers; limited real-world deployment
AI-supported intervention [108]	Reading and numeracy enhancement in intellectual and learning disabilities	10 AI-based sessions, Arabic-speaking students	Reading ↑ from 80.5 → 88.2; Numeracy ↑ from 78.7 → 86.4	Statistically significant gains; high teacher satisfaction; sustained improvement	Small-scale trial; limited long-term evaluation
AI-driven educational interventions [109]	Dyslexia, dyscalculia, dysgraphia skill enhancement	Multiple intervention studies reviewed	Moderate–large effect sizes	AI-based remediation superior to static drills and conventional interventions	Heterogeneity in intervention design; lack of longitudinal follow-up

The table outlines how ML, DL, EEG-based systems, and NLP tools have been applied to detect dyslexia, dysgraphia, dyscalculia, auditory processing disorders, and language-processing difficulties. Reported findings include improved diagnostic accuracy, faster screening, and enhanced identification of subtle cognitive and linguistic patterns. Common limitations include small or homogeneous samples, limited real-world validation, lack of cross-linguistic generalisability, and early-stage or laboratory-only evaluations

AI artificial intelligence, ML machine learning, DL deep learning, CNN convolutional neural network, RNN recurrent neural network, NLP natural language processing, EEG electroencephalography, ABR auditory brainstem response, APD auditory processing disorder, CAPD central auditory processing disorder, LPD language processing disorder, ABR auditory brainstem response

Motor disorders

Motor developmental milestones are another crucial component affected with neurodevelopmental disorders. Disorders such as cerebral palsy can lead to serious alterations in movement, posture and coordination, contributing to a decline in their quality of life. According to the DSM-5, motor disorders are one of the core categories within neurodevelopmental disorders that cause functional impairment. This includes stereotypical movement disorder (SMMs), developmental coordination disorder and tic disorders, including Tourette’s Syndrome [110].

Detection, diagnosis and classification

Early identification of motor neurodevelopmental disorders is essential, as brain plasticity is greatest during infancy. Recent advances in ML applied to infant movement and pose data have shown promise for the early detection of cerebral palsy. The BabyPose dataset, which includes recordings from 16 preterm infants, was developed to support pose-estimation research aimed at early motor assessment [111]. In a separate study, McCay et al. proposed a pose-based feature-fusion and classification framework that combined spatio-temporal and frequency–domain motion features from infant video data, achieving near-perfect predictive performance (AUC ≈ 0.99) for early cerebral palsy prediction [112]. Raghuram et al. also found that vertical velocity and movement quantity were key predictors in 2D video analysis, reporting 80% specificity and 93% negative predictive value for cerebral palsy detection [113].

SMMs, present in about half of individuals with ASD, were analysed using OpenPose skeletal extraction, region-based CNNs, and 3D-CNNs, achieving 92.5% recall and 66.8% precision with r = 0.80–0.88 to manual scores [114]. Lightweight neural networks applied to hand-flapping videos reached an F1-score of 84%, precision of 89.6%, and recall of 80.4% [115]. Accelerometer-based studies further showed that learned deep features outperform handcrafted features in detecting SMMs [116].

In developmental coordination disorder, an RCT showed that AI-supported handwriting training significantly improved all Minnesota Handwriting Assessment domains (p < 0.001; many effects large), whereas the control group showed minimal change (significant improvement only in alignment) [117]. DL models applied to SensoGrip smart-pen data achieved low error rates and strong agreement with expert assessments [118].

For tic disorders, Random Forest and DNN models achieved F1-scores of 82% and 79.5%, respectively, with overall accuracies around 88% [119]. A CNN model applied to MRI features improved diagnostic accuracy and reconstruction efficiency in children with tic disorder [120]. The wearable TSBand, integrating motion and vital-sign sensors, was endorsed by approximately 76% of caregivers for its usefulness in providing pre-tic alerts [121]. Lesion-network mapping further linked tics to distributed brain circuits centered on the basal ganglia, caudate, and globus pallidus externus, with negative connectivity to the precuneus rather than isolated regional damage [122].

A predictive model using olfactory and cognitive data achieved 100% accuracy, precision, and recall across four ML algorithms; the decision tree showed a Matthews Correlation Coefficient of 1. F1 scores were 100%, 91.7%, and 90% for the decision tree, MLP, and SGD classifiers [123].

Assessing gross motor function

The Gross Motor Function Measure (GMFM) evaluates motor abilities in children with cerebral palsy across five domains: lying and rolling, crawling and kneeling, sitting, standing, and walking–running–jumping. The original GMFM includes 88 items, while the revised GMFM-66 uses 66 items on an interval scale and takes 45–60 min to administer, which can challenge younger children’s attention spans [124].

To improve efficiency, self-learning algorithms, random forest, SVMs, and feed-forward neural networks were used to create a reduced GMFM-66 (rGMFM-66). Validation in 1,217 assessments showed excellent agreement with the full version, with intraclass correlation coefficients of 0.997 for single scores and 0.993 for yearly change detection [125]. The rGMFM-66 required only 34.5 items on average, produced mean absolute errors of about 1 point, and achieved 92.3–98.1% sensitivity and specificity for detecting clinically relevant change [125]. In 1352 children, raw score correlations reached r = 0.99 (p < 0.001), with individual effect size correlations of r = 0.84. Effect sizes during active training were nearly identical for the full and reduced tests (0.64 vs. 0.63), and diagnostic accuracy for detecting improvement or deterioration reached AUCs of 0.90 and 0.95, correctly classifying 85–95% of cases [126].

A complementary Medical Device Score Calculator (MDSC) estimated GMFM-66 scores from assistive device usage when direct testing was unavailable. In 1581 patients (mean age 8.1 years), a random forest model achieved a concordance correlation coefficient of 0.75 and mean absolute error of 7.74 points, supporting group-level but not individual assessment [127].

Gait analysis

AI models for gait and motion analysis are cost- and time-efficient, reducing the need for direct measurement of movement parameters. Pattern recognition in motion and gait datasets enables diagnosis and prediction of abnormal gait types typical of neurodevelopmental disorders such as cerebral palsy. Tools like Qualisys infrared cameras, inertial measurement unit sensors, and optoelectronic cameras are commonly used, though they face challenges such as complex setup, post-processing demands, accuracy loss over time, calibration needs, and high cost [128–131].

The recent advances in AI and ML have improved the assessment of gait in children with motor developmental disorders. One-dimensional CNNs trained with cross-validation have been shown to predict lower-limb joint moments in children with cerebral palsy with high accuracy [132]. Building on this work, a colour-coded classification system was introduced to visualise model performance, categorising predictions as green, yellow, or red based on error magnitude. Within this framework, the hip joint achieved the highest proportion of green classifications (84%), and significant differences in normalised root mean square error (nRMSE) were observed among all label groups across joints (p < 0.001) [133]. As prediction quality declined, kinematic parameters such as maximum angle, range of motion, and mean angles also decreased, while gait profile scores increased from 13.3° to 16.4° for the hip, 13.2° to 15.0° for the knee, and 13.1° to 15.0° for the ankle. Lower-quality predictions corresponded to greater gait abnormalities, suggesting that this colour-coded CNN framework provides a clear and clinically interpretable approach for reviewing AI-derived gait data [133].

For predicting patient’s response to treatment strategies

Few studies have used AI to analyse patients' responses to treatment strategies for motor developmental disorders. However, a recent study set out a protocol to investigate the prediction of treatment responses to a rehabilitation programme for children with cerebral palsy. The devised radiomics strategy includes the Gross Motor Function Classification System (GMFCS), the Assisting Hand Assessment (AHA), and the Manual Ability Classification System (MACS). The study aimed to record changes in motor scale values after three months. These were then compared to the baseline values measured using the Peabody Developmental Motor Scales (PDMS-2) [134]. Furthermore, researchers have utilised four DL algorithm models to support the control of exoskeletons used in rehabilitation for cerebral palsy and produces gait prediction and personalised trajectories. A 200 time-step trajectory was measured with the lowest mean absolute error (MAE) in the long-short-term memory (LSTM) and the CNN model [135]. Figure 6 illustrates the key strengths of the application of AI in motor disorders.

Fig. 6.

Overview of how AI supports the detection and assessment of motor disorders. AI models assist diagnosis by analysing movement patterns, facial features, and video-based motor behaviours, enabling automated classification in conditions such as cerebral palsy, Tourette syndrome, and developmental coordination disorder. ML systems also help evaluate gross motor function using reduced-item assessments with high accuracy. In addition, neural network–based gait analysis tools provide objective, clinician-friendly interpretations of walking patterns and motor performance. (Created with BioRender.com). AI artificial intelligence, ML machine learning, DNN deep neural network, SVM Support Vector Machine, MLP Multilayer Perceptron, CNN convolutional neural network

Critical appraisal: strengths, limitations, and research gaps of AI applications in motor developmental disorders

Although AI models have enhanced the detection and monitoring of motor neurodevelopmental disorders, current studies were conducted on small and homogeneous groups of participants, which limits the reproducibility of the results. Automated movement-analysis systems for cerebral palsy are accurate on their own video datasets but have not been tested in different clinical settings or recording conditions where variables such as lighting, posture and video length can be variable [136]. Recent AI models that analyse body movement patterns indicate a significant decrease in accuracy when tested on different sites, primarily due to the lack of shared labelling or recording standards. This underscores the importance of standard data collection and annotation procedures across studies [137].

Computer-vision and wearable devices can be used to objectively track repetitive or tic-like movements, but the available approaches rely on short manually labelled video sequences, limiting continuous, real-time tracking. DL models validated against laboratory data are unlikely to take into account factors such as medication use or comorbid conditions, limiting their ability to apply fairly and accurately to broader Tourette or autism populations [114]. Similarly, wearable sensor studies show feasibility of continuous detection of tics; however, most of these are based on closed algorithms that do not allow for independent verification [138].

In the context of gross motor function assessment, ML models have demonstrated that reduced versions of the GMFM-66 test have a similar level of accuracy to the full version and a shorter testing time. However, these studies have been performed mainly retrospectively and in specialist hospital settings rather than in community clinics [125]. Similarly, AI-based gait analysis systems can be used to estimate joint angles and movement patterns from wearable sensors, but their accuracy is sensitive to factors such as sensor calibration and walking speed, indicating that results can vary with hardware or data-processing differences [139].

Although AI-assisted rehabilitation devices, such as exoskeleton motion prediction, have proven to be technically feasible, little evidence exists that they produce sustained functional gains or improvements in daily life. The majority of studies are single-centre and are concerned with model accuracy rather than patient outcome or ease of use of the technology by therapists [135]. Future studies should involve multi-centre collaborations, open benchmarking and patient-centred measures to test if AI really does increase independence, mobility and overall participation. Table 4 discusses the models and datasets used, as well as the key outcomes of AI in motor disorders.

Table 4.

Summary of key AI models, datasets, and reported outcomes in motor developmental disorders

AI model/system	Application domain	Dataset/participants	Performance metrics (as reported)	Key findings/outcomes	Main limitations (as stated)n
Multilayer Perceptron (MLP) [111]	Detection of cerebral palsy in preterm infants	16 preterm infants	Accuracy 84% (outperformed K-Star, Naïve Bayes, Random Tree, SVM)	MLP outperformed other models for early CP classification	Very small sample; limited generalisation
Skeleton-based clustering (FFT-Joint Displacement) [112]	CP detection via skeleton movement segmentation	MINI-RGBD dataset	100% classification accuracy	Effective skeleton-based segmentation for motor disorder detection	Tested on specific dataset only; needs external validation
2D movement feature analysis [113]	CP detection in preterm infants	2D video movement data	Specificity 80%; PPV 26%; NPV 93%	Vertical velocity and motion quantity were key features	Limited sample size; low PPV reduces predictive utility
OpenPose + Region-CNN + 3D-CNN [114]	Detection of stereotyped movements (SMMs) in ASD	220 train/21 test children (ASDPose 580 h dataset)	Recall 92.53%; Precision 66.82%; r = 0.80–0.88 (vs manual)	Accurate SMM detection; corrected > 50% false positives on re-annotation	Moderate precision; dataset specific; ASD-only context
Lightweight NNs (× 5) [115]	SMM classification (flapping vs control)	Video dataset of children	F1 = 84; Precision 89.6%; Recall 80.4%	Effective video-based detection using compact models	Small datasets; limited generalisation
Accelerometer feature models [116]	Wearable SMM detection	Not specified	Not quantified (“high-quality requires handcrafted or learned features”)	Established feasibility for mobile monitoring	Requires hand-crafted features; not real-time
AI-supported occupational therapy [117]	Developmental coordination disorder (handwriting therapy)	Randomised controlled trial	p < 0.001 across all subdomains; Cohen’s d > 0.8	Significant improvement on MHA subdomains vs controls	Short-term study; specific to handwriting
DL on SensoGrip pen data [118]	Predicting motor handwriting difficulty scores	SensoGrip data + expert evaluations	Low RMSE; strong agreement with experts	Reliable AI prediction of handwriting metrics	Limited sample and domain scope
Random Forest + DNN [119]	Tic detection in Tourette syndrome videos	Video dataset of patients	RF: F1 82%; Accuracy 88.4%/DNN: F1 79.5%; Accuracy 88.5%	Automated tic classification using facial landmarks and videos	Dataset specific; no clinical deployment
Hybrid unsupervised + supervised framework [120]	Tic detection (binary/multiclass)	Video datasets	Precision > 77%; Recall > 77%	Effective multi-task tic classification	Moderate performance; not validated clinically
TSBand wearable [121]	Wearable tic attack alert system	User survey	76% participants favoured device	Shows user acceptance of wearable monitoring	Prototype; not tested clinically
Olfactory + Cognitive Model [123]	Tourette syndrome prediction	Clinical cohort (not stated)	Decision Tree: Accuracy 100%; MCC 1/MLP: F1 91.66%/SGD: F1 90%	Demonstrated objective TS diagnosis across 4 ML models	Needs independent validation and larger samples
rGMFM-66 (Random Forest, SVM, NNs) [125]	Reduced Gross Motor Function Measure for CP	1 217 assessments	ICC 0.997 (single)/0.993 (change); Sens./Spec. 92.3–98.1%	Shortened test time (34.5 items vs 66) with maintained accuracy	Retrospective hospital data only
rGMFM-66 Validation [126]	Further validation of rGMFM-66 in CP	1 352 children with CP	r = 0.99 (p < 0.001); AUC 0.90–0.95; Correct classification 85–95%	Confirmed high agreement and diagnostic accuracy	Conducted in specialist centres only
Medical Device Score Calculator (MDSC) [127]	Estimating GMFM-66 scores from device data	1 581 patients (mean age 8.1 y)	CCC 0.75; MAE 7.74 points	Enables group-level motor function estimation without direct testing	Not suitable for individual use
1D Neural Network Gait Model [132]	AI-based gait analysis in motor disorders	Children with motor disorders	Hip moment green 84%; p < 0.001 (nRMSE diffs); Gait profile scores ↑ with severity	Colour-coded classification reflects severity and accuracy	Requires sensor calibration; site-specific
Radiomics Prediction Protocol [134]	Predicting rehabilitation response in CP	Protocol design study	(no results yet)	Planned use of GMFCS, AHA, MACS to predict motor scale change	Protocol only; pending results
Exoskeleton Trajectory DL Models (LSTM, CNN) [135]	Rehabilitation control and gait prediction in CP	Exoskeleton time-step data (200 steps)	Lowest MAE for LSTM and CNN	Enabled personalised trajectory prediction for therapy	Focused on model accuracy; no patient outcome data

The table outlines how ML, DL, wearable sensors, video-based analysis, and hybrid frameworks have been used to detect cerebral palsy, stereotyped movements, tic disorders, and motor coordination difficulties. Reported outcomes include improved early diagnosis, automated movement classification, enhanced handwriting assessment, and refined gait or motor-function scoring. Limitations relate mainly to small or highly specific datasets, limited clinical deployment, laboratory-only testing conditions, and prototype systems that require further validation

AI artificial intelligence, ML machine learning, DL deep learning, CP cerebral palsy, FFT fast fourier transform, PPV positive predictive value, NPV negative predictive value, ASD autism spectrum disorder, SMM stereotyped motor movement, NN neural network, RNN recurrent neural network, CNN convolutional neural network, DNN deep neural network, RF random forest, R-CNN region-based convolutional neural network, LSTM long short-term memory, MHA movement handwriting assessment, MDSC medical device score calculator, ICC intraclass correlation coefficient, Sens. sensitivity, Spec. specificity, AUC area under the curve, CCC concordance correlation coefficient, MAE mean absolute error, nRMSE normalised root mean square error, GMFM-66 Gross Motor Function Measure–66, GMFCS Gross Motor Function Classification System, AHA Assisting Hand Assessment, MACS Manual Ability Classification System

Communication/speech disorders

Individuals with neurodevelopmental disorders often experience a variety of communication and speech difficulties that can significantly impact their social interactions, learning and daily functioning. These include language delays and atypical prosody in ASD and childhood apraxia of speech (CAS), where the planning and coordination required for speech are disrupted [140, 141]. Other conditions include dysarthria, a motor speech disorder often seen in cerebral palsy; developmental language disorder (DLD), where children struggle to acquire language without an apparent neurological cause; and aphasia [142–144]. Some individuals may experience conditions such as selective mutism, echolalia (repetitive speech), cluttering (rapid, disorganised speech) or mixed expressive-receptive language disorder in severe cases.

In recent years, advances in AI have started to change how we understand, diagnose, and treat communication and speech disorders. These tools are helping in early diagnosis, effective progression monitoring, and enhanced therapy [145, 146]. These technologies are also helping to close gaps in access to care, increase diagnostic precision, and enable more personalised treatment strategies.

Early detection and classification

Early detection is crucial for the effective management of communication and speech disorders, particularly in neurodevelopmental conditions where language decline is often one of the first noticeable symptoms [147, 148]. A delay in diagnosis usually means limited potential impact of the therapeutic intervention. AI-powered speech analysis has proven to be a promising solution in this regard. Using ML and NLP, researchers can analyse acoustic, phonetic, and linguistic patterns that are often overlooked in clinical evaluations of conditions such as ASD, CAS, dysarthria, or stuttering. AI can also classify language impairments by extracting speech biomarkers such as pauses, articulation errors, and syntactic complexity. This allows for the more objective and earlier detection of aphasia. Large speech data can be used to train algorithms to detect early signs of neurodevelopmental speech disorders even before clinical symptoms arise [145, 148]. This improves diagnostic accuracy and enables continuous monitoring, which is important for tracking the progress of neurodevelopmental disorders [147, 148].

Comparative analysis shows that AI models frequently outperform traditional clinical-led assessment in early-stage detection and are even better when combined with the existing human-led assessment. This is due to the ability of AI to analyse and process thousands of acoustic and linguistic features simultaneously, while humans are limited by how they perceive these cues [142, 149]. Children with ASD and speech disorders show atypical intonation patterns, pitch and rhythm, or delayed babbling, which may go unnoticed in standard screening tests. AI models trained with speech acoustics and language patterns can detect sudden deviations. For instance, it has been demonstrated that AI-based speech recognition of child–examiner dialogue can distinguish between ASD and neurotypical development with almost 88% accuracy, using features such as prosody, pause patterns, and spectral qualities [150]. Similarly, home-recorded speech collected through the Guess What? mobile app by applying random forest models, CNNs, and fine-tuned Wav2vec 2.0 models were analysed [145]. In natural, real-world settings, their best-performing model was able to differentiate between ASD and typical speech with up to 79% accuracy [145]. This demonstrates the potential for AI to be used for diagnosis outside of clinical environments, making way for earlier and more accessible ASD screening.

AI has also been successfully applied to the detection and severity classification of conditions such as dysarthria. Pre-trained Wav2vec 2.0 embeddings were applied to dysarthric speech samples from the Universal Access Speech corpus. The severity classification of dysarthria improved by 10% compared to traditional Mel-frequency cepstral coefficients-based systems [142]. Building on this, a CNN model based on continuous wavelet transform features to classify dysarthria severity using the UA-Speech and TORGO datasets was also developed [143]. This system can accurately detect multilevel severity without relying on manual feature extraction [143]. Furthermore, the introduction of attention-based learning models to predict listener effort and classify speech deterioration highlights how interpretable AI models could support clinicians in both diagnosis and progress monitoring [151].

Stuttering or speech disfluency is another neurodevelopmental communication disorder in which AI has made progress. While the detection of stuttering used to be based on the subjective judgement of clinicians, AI now offers more objective and effective solutions. An acoustic SVM model has been developed that can distinguish between stuttering individuals and neurotypical speakers with 88% accuracy across various speech tasks and age groups [152]. Improvements in training approaches, such as balancing the data and augmenting examples, have led to better detection of stuttering. For example, F1-scores of 0.97 and 0.96 were achieved using CNN and ConvLSTM models, respectively, on the SEP-28 K dataset [153]. Similarly, fine-tuned Wav2vec 2.0 models and multi-task learning improved F1 scores for event-specific stuttering detection by up to 27% [153].

Some AI-focused research has been conducted on CAS. One approach uses a DNN-based lexical stress classification tool to evaluate stress contrast production in polysyllabic words. This system achieved over 80% agreement with expert speech and language therapists for weak-strong patterns, which are more complex [140]. Another important innovation is the Tabby Talks system, which was designed for remote CAS therapy [140]. It integrates modules for voice activity detection, pronunciation verification, and stress analysis, enabling it to identify errors in motor planning or speech execution, although quantitative performance data was not fully reported [141].

In addition, developmental language disorder and mixed receptive-expressive language impairment have been explored through acoustic markers such as voice patterns, timing and rhythm. Although large-scale ML detection frameworks for these disorders are still in their infancy, preliminary findings suggest that ML could be a valuable tool for identifying subtle linguistic deficits that might otherwise remain undetected during early childhood screening [144].

Overall, AI has proven to be a versatile tool for identifying a broad range of neurodevelopmental communication disorders. From ASD and dysarthria to stuttering and CAS, as well as in the emerging field of DLD, AI systems are able to identify speech patterns that human perception alone cannot reliably capture. AI has the potential to transform early detection and intervention by providing earlier, more objective and more scalable screening, ensuring that children with speech and communication challenges receive timely, tailored support.

Improving speech and language therapy

The type of therapy provided is crucial for managing and recovering from a neurodevelopmental communication disorder. Traditional speech and language therapy relies heavily on in-person clinician direction, which can be costly and time-consuming, and is not always accessible. AI-driven therapies are now being used to augment and personalise therapy, making it more effective, adaptive and interactive.

For instance, real-time feedback, adaptive difficulty levels, and gamified therapy sessions have been demonstrated to enhance aphasia rehabilitation engagement when AI-based tools are employed, replacing the fixed-pace, one-size-fits-all approach with personalised, engaging sessions [149]. Similarly, AI-assisted platforms can treat speech sound disorders by using automated systems to track articulation patterns and provide corrective prompts [146]. In the case of dysarthria, DL models that use continuous wavelet transform–based CNN can objectively identify changes in speech features such as articulation, pitch, and loudness. These quantitative measures allow clinicians to adjust therapy plans based on measurable data rather than subjective assessment [155].

DL systems that integrate acoustic and visual cues have been shown to assess speech motor disorders more accurately than manual evaluations, reducing variability and improving diagnostic confidence [156]. AI contributions go beyond intervention planning. LLM-driven platforms can automatically generate personalised therapy plans by synthesising clinical guidelines with patient-specific data, significantly reducing the administrative workload while maintaining alignment with best practices [157]. In aphasia treatment, systems that analyse speech quality parameters can recommend activities whose difficulty adapts as performance improves. This offers a level of responsiveness that is rarely possible with static, workbook-based approaches [158].

Although these AI-supported tools cannot replace therapists, they can extend their reach and improve care by enabling patients to practise exercises independently while maintaining some level of professional oversight. They also generate a large amount of progress data, helping clinicians to better understand which strategies work for specific individuals.

Enhancing the use of augmentative and alternative communication interventions for people with communication disorders

Augmentative and alternative communication (AAC) devices are important tools that enable people with severe speech and language impairments to communicate. However, traditional AAC systems typically require labour-intensive customisation and lack the ability to adapt dynamically to context, making personalisation challenging. However, advances in AI have begun to transform AAC systems into more intuitive, predictive and context-aware solutions.

AI-driven AAC prototypes that incorporate features such as instant phrase suggestions, visual verification of intended words, and context-aware correction of user input have been developed for people with aphasia [159]. These systems can anticipate user needs and offer communication shortcuts, thereby reducing the cognitive load required to use the device. Similarly, a study emphasised that integrating AI with AAC can improve the prediction of user intent, facilitate quicker vocabulary selection and adapt based on environmental cues such as ongoing conversations or locations [160]. The system can deliver contextually appropriate, emotion-sensitive suggestions by recognising facial expressions, head movements and environmental changes, thereby creating a more natural and responsive communication experience.

Predictive algorithms have also been shown to enhance the effectiveness of generated messages by learning from previous communication patterns, thereby enabling users to express themselves more efficiently [161]. Traditional AAC often relies on manual scanning or categorised symbol searches, but AI prediction models substantially reduce selection time while maintaining accuracy. Similarly, incorporating AI language models into AAC systems can improve grammatical accuracy and speed up message formulation without limiting user autonomy [162]. This does not replace traditional symbol-based communication, but rather strengthens it by providing a more fluid bridge between symbol selection and natural-sounding speech output. Furthermore, multimodal AI frameworks have recently integrated wearable physiological sensors, computer vision and NLP into an AI-driven AAC framework. This framework has been shown to predict vocabulary and adapt to the user’s emotional and physiological states [163]. Such systems can predict vocabulary needs while dynamically adjusting to fluctuations in mood, stress levels and physical responses, creating a communication experience that is personalised and empathetic.

AI-enhanced AAC devices also restore a sense of autonomy and spontaneity, enabling users to engage in more natural conversations and improving their quality of life. Although these technologies are still emerging, they hold immense potential for improving the lives of people with severe communication impairments.

Critical appraisal: strengths, limitations, and research gaps of AI applications in communication disorders

AI has shown substantial promise in advancing the detection, classification, and intervention of speech and language disorders across neurodevelopmental and acquired conditions, including ASD, dysarthria, stuttering, CAS, aphasia, DLD, and AAC systems [142, 148, 149, 160]. DL architectures, such as convolutional, recurrent, and transformer-based networks, have demonstrated notable accuracy in identifying disorder-specific prosodic, articulatory, and linguistic anomalies [146, 150, 155]. These systems capture subtle acoustic and temporal features that are often difficult for clinicians to quantify, offering scalable, objective tools for assessment and therapy [147, 159].

Despite these advancements, critical methodological and translational limitations persist. Dataset diversity and representativeness remain major barriers. Most studies rely on small, homogeneous, English-language corpora collected in controlled environments, such as UA-Speech or TORGO, limiting generalisability and ecological validity [151, 155]. Multilingual, spontaneous, and conversational speech data remain underrepresented, reducing the models’ applicability to real-world contexts [150, 152]. Furthermore, longitudinal and developmental analyses are scarce, hindering the ability to model change over time in conditions such as ASD, CAS, or DLD [140, 144].

Another key limitation concerns interpretability and transparency. DL models often achieve high accuracy but operate as “black boxes,” offering limited insight into the linguistic or acoustic features driving their predictions [148, 149]. This lack of explainability impedes clinical trust and integration into evidence-based practice. The implementation of XAI approaches such as attention mapping and feature attribution could improve interpretability and align model reasoning with clinical frameworks.

Moreover, validation beyond research contexts remains limited. Many models perform well on curated datasets but lack independent evaluation in clinical or community settings [142, 157]. The predominant focus on unimodal audio data also neglects potentially informative multimodal cues such as articulatory movement, facial dynamics, and gesture patterns, which could enhance diagnostic accuracy [141, 151].

Ethical and user-centred considerations are increasingly pressing, particularly for adaptive AAC systems that infer user intent or emotion. While such systems improve communication efficiency, they raise concerns about autonomy, over-personalisation, and user agency [161]. Similarly, digital speech analysis and intervention tools for disorders such as aphasia or dysarthria raise privacy and data protection challenges that remain insufficiently addressed [157]. Transparency around methods also remains a challenge. Studies often describe feature extraction, model tuning, and data preprocessing inconsistently, making it difficult to reproduce findings or assess models for clinical and regulatory use [144, 148].

Several research gaps also remain that limit the clinical translation and scalability of AI in speech and language disorder management. First, the lack of large-scale, demographically and linguistically diverse datasets constrains model generalisability, particularly across underrepresented populations and non-English speakers [143, 150]. The absence of longitudinal and developmental studies also impedes understanding of how AI can track progression or therapeutic response over time in disorders such as DLD, and CAS [140, 144].

Furthermore, interpretability remains a major challenge for clinical adoption. Models that function as black boxes can perform well but offer little insight into how their predictions are made, which makes it difficult for clinicians to trust and apply them in practice [148, 149]. From a translational standpoint, a clear gap persists between experimental validation and real-world clinical use, since only a handful of studies have implemented these tools in live healthcare or community settings [142, 157]. The ethical dimensions of adaptive AAC and continuous monitoring systems, particularly issues around autonomy, consent, and data privacy, are also not yet fully theorised or regulated [161]. In addition, inconsistent methodological reporting makes replication and regulatory evaluation difficult, highlighting the need for more transparent and standardised research practices [144, 148]. Addressing these challenges will require collaboration among clinicians, computational scientists, and ethicists to develop AI systems that are not only technically robust but also clinically interpretable, equitable, and ethically grounded. Table 5 summarises the AI models and datasets used, and the key outcomes of AI in relation to communication disorders.

Table 5.

Overview of AI models, datasets, and reported outcomes in communication disorders

AI model/system	Application domain	Dataset/participants	Performance metrics (as reported)	Key findings/outcomes	Main limitations (as stated)
DNN-based lexical stress classifier [140]	Childhood Apraxia of Speech (CAS) stress pattern evaluation	Speech samples from children with CAS	> 80% agreement with expert therapists (weak–strong stress)	Enabled objective stress contrast classification	Limited sample size; only partial performance data reported
Tabby Talks remote CAS therapy system [141]	Remote AI-supported CAS intervention	Clinical and home-based pilot users	Qualitative outcome (no quantitative accuracy reported)	Integrated modules for pronunciation and motor planning error detection	Lacked quantitative validation; small pilot scope
Pre-trained Wav2vec 2.0 embeddings [142]	Dysarthria severity classification	Universal Access Speech (UA-Speech) corpus	10% improvement over MFCC-based baseline	Automated dysarthria severity grading with improved accuracy	Limited dataset diversity; English-only samples
CNN using continuous wavelet transform features [143]	Dysarthria detection and severity estimation	UA-Speech and TORGO datasets	High accuracy across multilevel severity classes (exact % not stated)	Eliminated need for manual feature extraction	Dataset small and non-representative; lacks multilingual data
Transformer-based speech analysis [144]	Developmental Language Disorder (DLD) detection	Child speech recordings; preliminary dataset	Not specified (pilot-level results)	Demonstrated potential of acoustic markers for DLD screening	Early-stage research; lacks large-scale validation
Random Forest, CNN, fine-tuned Wav2vec 2.0 [145]	Early detection of ASD through speech analysis	Guess What? home-recorded child–examiner dialogues	Up to 79% accuracy in differentiating ASD from neurotypical speech	Demonstrated feasibility of AI-based ASD screening in real-world environments	Small dataset; limited linguistic and cultural diversity; real-world noise sensitivity
Adaptive AI-driven aphasia rehabilitation platform [149]	Aphasia therapy and progress monitoring	Adult aphasia patients (controlled clinical study)	Improved engagement and rehabilitation outcomes qualitatively	Real-time feedback and adaptive difficulty improved therapy efficiency	Small sample; limited longitudinal outcome data
Deep learning speech classifier [150]	Speech biomarker detection for ASD	Clinical dialogue corpus of children with ASD and controls	88% classification accuracy	Identified atypical prosody, rhythm, and pause patterns predictive of ASD	Controlled clinical data; lacks longitudinal validation
Attention-based regression model [151]	Speech deterioration prediction and listener effort estimation	Clinical speech recordings from dysarthric patients	Correlation coefficients improved over baseline models	Introduced interpretable AI for assessing deterioration	Requires multimodal validation and clinical deployment
SVM, CNN, and ConvLSTM models [152, 153]	Stuttering (speech disfluency) detection	SEP-28 K stuttering speech dataset	F1 = 0.97 (CNN); F1 = 0.96 (ConvLSTM)	Reliable automatic stutter detection across tasks and age groups	Dependence on balanced, annotated data; limited spontaneous speech inclusion
Fine-tuned Wav2vec 2.0 with multi-task learning [154]	Event-specific stuttering detection	SEP-28 K	Up to 27% improvement in F1 over baseline	Enhanced detection of specific stuttering events	Model generalisation across languages not tested
LLM-based therapy plan generator [157]	Automated speech therapy planning	Clinical datasets and guideline synthesis	Not quantitatively reported	Reduced clinician workload; improved therapy personalisation	Limited explainability; no clinical validation data
AI-augmented AAC prototype [159]	AAC enhancement for aphasia	Individuals with acquired aphasia	Not specified; qualitative usability evaluation	Context-aware phrase prediction and visual verification enhanced communication	Limited sample size; controlled testing
Predictive AAC with intent and emotion detection [160, 163]	AAC systems for severe communication disorders	Mixed participants using AAC devices	Qualitative user satisfaction improvements	Contextual and emotion-sensitive vocabulary prediction enhanced autonomy	Ethical issues around autonomy and over-personalisation; data privacy risks

The table summarises how DL, transformer-based systems, traditional ML, and AI-supported communication devices are used for detecting speech impairments, assessing disorder severity, supporting therapy, and enhancing augmentative and alternative communication (AAC). Reported outcomes include improved classification of dysarthria, apraxia, developmental language disorders, autism-related speech features, and stuttering patterns. Limitations commonly involve small or homogeneous datasets, early-stage or qualitative findings, limited multilingual testing, and a lack of large-scale clinical validation

AI artificial intelligence, ASD autism spectrum disorder, CNN convolutional neural network, DLD developmental language disorder, DL deep learning, DNN deep neural network, ML machine learning, SVM Support Vector Machine, AAC augmentative and alternative communication, BCI brain–computer interface, LLM large language model, MFCC mel-frequency cepstral coefficient, UA-Speech universal access speech, CAS childhood apraxia of speech

Latest advancements: combining AI with robotics for the management and research of neurodevelopmental disorders

Recent advances in AI and robotics have significantly transformed the management of neurodevelopmental disorders, offering novel opportunities for both diagnosis and therapy. AI systems, particularly those using ML and LLMs, have shown promise in automating aspects of diagnosis and providing personalised interventions [163–165]. Robotics, on the other hand, have demonstrated strong results in enhancing engagement and supporting daily living skills. Socially assistive robots like LOLA2 have successfully reinforced activities of daily living in individuals with neurodevelopmental disorders, providing real-time monitoring of user performance and receiving positive feedback from therapists and families [166]. Similarly, robot-assisted therapy has been effective in improving attention and imitation skills in children with autism and intellectual disabilities, particularly by leveraging DL to estimate visual focus during therapy sessions [167].

While both AI and robotics have demonstrated independent benefits, recent studies suggest that integrating these technologies produces superior outcomes by combining adaptive cognitive interaction with embodied physical presence. AI-only conversational agents such as ChatGPT can foster cognitive development and language skills by engaging children in tasks like turn-taking, sentence formation, and problem solving [168].

However, comparative studies indicate that children engaging with humanoid robots powered by the same AI show higher cognitive focus, improved social behaviors, and greater engagement than those interacting with AI through screens alone [169, 170]. This advantage arises from the robots’ ability to deliver non-verbal cues, such as gestures, gaze, and facial expressions, which provide richer sensory input and better mimic real-life interactions [165].

In therapeutic contexts of ASDs, humanoid robots integrated with AI, such as Pepper linked with ChatGPT, have facilitated emotion recognition, imaginative play, and problem solving in children with low-to-medium functioning ASD, providing adaptive interactions that adjust to each child’s responses [170]. DL methods incorporated into these robots allow for real-time monitoring of engagement, enabling therapists to tailor interventions dynamically and objectively [167, 171].

For ADHD, AI robotics integration has also shown measurable therapeutic benefit. ChatGPT-4o embedded within socially assistive robots has been used to deliver adaptive cognitive exercises and behavioral prompts, with eight-week trials reporting increases in attention span by approximately 28% and reductions in hyperactivity by about 22% [84]. These platforms also address scalability challenges in traditional therapy by facilitating multi-user interactions and reducing the need for continuous therapist supervision, a critical factor in resource-limited settings.

Beyond ASD and ADHD, AI robotics frameworks have demonstrated value for children with intellectual disabilities and broader learning disorders. ML models using multimodal data achieved engagement detection accuracies above 90% in special education contexts, allowing robots to adapt educational scenarios to maintain student participation [171]. This adaptability is especially valuable for children with dyslexia or other learning difficulties, where sustained motivation and individualised reinforcement are essential for effective learning.

Overall, evidence across these studies indicates that while AI and robotics each contribute significant advances in managing neurodevelopmental disorders, their combination provides a more immersive, responsive, and effective therapeutic environment. By merging the adaptive cognition of AI with the physical and social presence of robots, these integrated systems enhance engagement, accelerate skill acquisition, and support scalable interventions tailored to the diverse needs of individuals with neurodevelopmental disorders [84, 166, 170]. Table 6 summarises the outcomes of combining robotics and AI in the management and research of neurodevelopmental disorders.

Table 6.

Summary of outcomes from AI-robotic interventions used in neurodevelopmental disorders

Author, country (year)	Neurodevelopmental disorder type	AI model type used	Robotic system	Main outcomes
López Requena et al., Spain (2022) [169]	General neurodevelopmental disorders affecting ADLs	DL (3D CNN)	Humanoid robot	Compared AI-only screen-based interaction vs humanoid robot interaction; the robot-assisted group showed higher cognitive focus, improved social behaviour, and greater engagement
Holeva et al., Greece (2024) [165]	ASD, communication, learning, and motor disorders	DL (OpenPose, OpenFace)	Child-sized humanoid robot (NAO)	Robot-assisted intervention improved psychosocial skills compared with human-only intervention
Di Nuovo et al., Italy (2018) [167]	ASD and learning disabilities	ML (KNN); DL (Faster R-CNN, MTCNN)	Humanoid robot (SoftBank Robotics NAO v4)	Robot-assisted therapy improved attention and imitation skills; DL models estimated visual focus in real time to personalise therapy
Papakostas et al., Greece (2021) [171]	Learning disabilities	ML (Random Forest, SVM); Multimodal DL	Humanoid robot	Achieved > 90% engagement detection accuracy, enabling adaptive educational content
Berrezueta-Guzman et al., Germany (2025) [84]	ADHD	LLM (ChatGPT-4o + reinforcement learning)	Socially assistive humanoid robot	Eight-week trials showed 28% increase in attention span and 22% reduction in hyperactivity; reduced therapist supervision needs
Nasri et al., Spain (2022) [166]	Motor and neurometabolic disorders	DL (multimodal ADL monitoring model)	Socially assistive robot (LOLA2)	Enhanced ADL performance through real-time monitoring and adaptive reinforcement; positive feedback from therapists and families

The table highlights how machine learning, DL, and LLMs are integrated into socially assistive and humanoid robots to support cognitive, social, behavioural, and motor development. Across studies, AI-enabled robots enhanced engagement, improved attention and communication skills, supported activities of daily living, and enabled personalised interaction through real-time behavioural monitoring. Reported limitations include small pilot samples, short trial durations, and limited clinical generalisation

AI artificial intelligence, ML machine learning, DL deep learning, ASD autism spectrum disorder, ADHD attention-deficit hyperactivity disorder, CNN convolutional neural network, ADL activities of daily living, SVM Support Vector Machine, LLM large language model, R-CNN region-based convolutional neural network, MTCNN multi-task cascaded convolutional neural network, KNN K-nearest neighbour

Discussion and perspectives

Incorporating model interpretability frameworks such as XAI

As AI technologies increasingly influence diagnostic and therapeutic pathways in neurodevelopmental disorders, the demand for interpretability has become a central concern. The opaque "black-box" nature of many ML models can undermine clinical trust and hinder adoption in sensitive medical domains such as paediatrics and psychiatry. XAI offers a crucial framework to bridge this gap by elucidating model decisions in transparent and clinically meaningful ways. By providing post hoc interpretations or inherently interpretable architectures, XAI enhances accountability and facilitates collaboration between AI systems and healthcare professionals [172, 173].

Moreover, the recent studies indicate that integrating XAI into neurodevelopmental assessment frameworks may improve the interpretability of multimodal data, including clinical, behavioral, and neurobiological features. Evidence from survey research suggests that XAI can aid diagnostic reasoning in complex neurodevelopmental profiles [174], while systematic reviews highlight its role in promoting transparency and adaptive support within real-world environments [53]. As research progresses, embedding XAI into clinical workflows will not only foster user confidence but also comply with emerging regulatory expectations around algorithmic transparency and accountability [174, 175].

Beyond developing XAI, researchers argue that the black‑box problem in neurodevelopmental AI can be mitigated by improving how systems are built and integrated into clinical practice. It is recommended to establish gold‑standard protocols that are understandable to practitioners and openly communicate algorithm limitations to overcome this issue [8]. The black‑box perception regarding intellectual and developmental disabilities stems partly from poor data quality and complexity; solving this may involve increasing dataset resolution, investing in computational capacity and developing sophisticated yet still interpretable models that will help bridge the gap between research and clinical implementation. These same studies also highlight practical measures such as using oversampling and other unbiased approaches to address small sample sizes and class imbalance. This can be further aided by adopting clinically oriented evaluation metrics like decision‑curve analysis rather than purely statistical indicators and building centralised data‑sharing infrastructures, with careful metadata annotation [176]. Training clinicians to interpret AI outputs and developing regulatory guidance will further ensure that AI remains a tool that augments rather than obscures clinical judgement [8].

In addition, real-world adoption remains constrained by practical implementation barriers, including the need for workforce training, formal regulatory validation, and sufficient outcome-level evidence to support reimbursement and routine clinical use.

Developing new and bigger datasets

Recent strategies in AI for neurodevelopmental disorders have shifted toward integrating multimodal datasets, combining clinical, behavioral, genomic, and neuroimaging information into unified repositories [8, 177]. This integrative approach allows models to capture the multifactorial nature of neurodevelopmental disorders, enabling predictions that reflect both biological mechanisms and behavioral presentations rather than isolated data streams.

However, the success of these initiatives remains constrained by fundamental issues in data quality and representativeness. Many existing datasets are small, lack diversity, or follow inconsistent standards, which limits the robustness and transferability of AI models across different clinical and demographic groups [53, 178]. Addressing these gaps requires not just incremental improvements but the deliberate construction of large, inclusive datasets spanning varied ages, ethnicities, and phenotypic profiles to ensure equitable and accurate model performance [53, 178].

Parallel to these data‑centric efforts, new frameworks such as federated learning are emerging to overcome institutional and privacy barriers to large‑scale data sharing [174, 179]. By enabling distributed model training without centralising sensitive patient information, these methods create opportunities to scale neurodevelopmental AI research collaboratively across institutions while maintaining stringent privacy protections.

Addressing ethical issues such as privacy, security, and transparency

Ethical considerations remain central to the integration of AI within neurodevelopmental care, where vulnerable paediatric populations face heightened risks. The primary concern involves ensuring fairness and preventing harm; biased algorithms derived from unrepresentative datasets have the potential to reinforce diagnostic inequities or produce unsafe treatment recommendations [173, 175].

Equally critical is the challenge of maintaining transparency and respecting patient autonomy. When AI models operate as opaque “black boxes,” clinicians and families are unable to scrutinise or contest their decisions, undermining informed consent and contravening the ethical principle of non‑maleficence [172, 179]. Embedding explainability measures into both system architecture and institutional oversight processes is therefore necessary to sustain trust and support ethical clinical decision making.

Beyond transparency and bias, safeguarding personal health information represents another priority. Paediatric neurodevelopmental data are often highly sensitive, and robust privacy protections must be in place to prevent misuse or unauthorised access. This highlights the need for dedicated ethical frameworks tailored to AI in mental health and neurodevelopment, providing guidance on responsible innovation and ensuring trust between families, clinicians, and researchers [8, 175].

Furthermore, legal frameworks such as the California Consumer Privacy Act, the Health Insurance Portability and Accountability Act, and the EU General Data Protection Regulation provide essential foundations for data protection, yet they often fail to address the complexities of large-scale, adaptive AI systems that process sensitive patient information [180]. These systems may inadvertently enable re-identification of anonymised data or expose confidential information, underscoring the need for continuous algorithmic auditing and compliance monitoring. The lack of harmonised international standards further amplifies vulnerabilities in data sharing and informed consent, especially in cross-border healthcare contexts. Effective mitigation requires privacy-preserving techniques such as differential privacy, federated learning, and homomorphic encryption, along with consistent accountability frameworks and transparency obligations throughout the AI lifecycle [180, 181].

Cost-effectiveness and access, especially in low-resourced settings

AI holds promise for addressing disparities in the diagnosis and treatment of neurodevelopmental disorders, particularly in low- and middle-income countries. In these regions, healthcare infrastructure is often underdeveloped, and access to specialists is limited. AI-based screening tools and mobile diagnostic platforms can extend care to underserved populations by automating assessments and supporting task-shifting strategies [182, 183].

However, ensuring cost-effectiveness and sustainability of these interventions remains a significant challenge. While initial implementation costs may be high, long-term benefits in efficiency, early detection, and resource allocation are increasingly being documented [184, 185]. Public–private partnerships, open-source AI solutions, and adaptive technologies tailored to local contexts may further improve scalability and equity [183, 186]. Bridging the digital divide through inclusive policy frameworks and infrastructure investment is essential to realise AI’s full potential in global neurodevelopmental care.

Limitations of study

Although a broad search strategy was used, studies published in languages other than English were excluded, possibly overlooking relevant findings. The field of AI in neurodevelopmental disorders is rapidly evolving, and many of the included studies involved small sample sizes or single‑site datasets. High‑dimensional data coupled with limited participant numbers can lead to model over‑fitting and poor generalisability to new populations. In addition, most studies relied on internal cross-validation rather than independent external validation, limiting confidence in the robustness and generalisability of reported performance metrics [18, 41]. Heterogeneity in data collection methods, varying diagnostic criteria and inconsistent reporting standards across studies further complicate comparisons and synthesis of results [8]. Algorithmic bias presents another limitation. Language models have been shown to associate neurodevelopmental terms with negative concepts, and AI systems trained on non‑representative datasets may perpetuate disparities.

Conclusion

AI applications are reshaping research and care for neurodevelopmental disorders by enabling early diagnosis, adaptive interventions and ongoing support. As these technologies mature and move closer to routine clinical integration, several strategic research priorities must be addressed to ensure they are safe, equitable and clinically meaningful.

Future Research Agenda;

• Increasing dataset diversity across age ranges, ethnicities, languages, socioeconomic contexts, and comorbidity profiles to improve fairness and generalisability.

• Establishing robust regulatory frameworks for clinical-grade AI, including standards for auditing, risk classification, model updates, and post-deployment surveillance.

• Advancing XAI methodologies to enhance transparency, clinician trust, and alignment with diagnostic reasoning.

• Assessing cost-effectiveness and long-term sustainability of AI-enabled screening, diagnostic and therapeutic pathways, especially in resource-limited settings.

• Conducting large-scale, multi-centre clinical trials with harmonised protocols to evaluate real-world performance, safety, and long-term developmental impact.

Advances span clinical decision support, personalised learning, motor rehabilitation and, increasingly, interactive therapies that combine AI with socially assistive robotics. Translating these innovations into practice requires attention to model transparency, bias reduction, data privacy and equitable access. Well‑designed, ethically grounded systems that augment clinical judgement could ultimately improve outcomes and quality of life for individuals living with neurodevelopmental disorders. A key question for the next decade is whether AI will primarily serve as a screening tool, a clinician-support system, or eventually evolve toward an independent diagnostic role.

Abbreviations

AI

Artificial intelligence

ML

Machine learning

DL

Deep learning

ASD

Autism spectrum disorder

ADHD

Attention-deficit hyperactivity disorder

ANNs

Artificial neural networks

MRI

Magnetic resonance imaging

DNN

Deep neural network

ICHD

International Classification of Headache Disorder

SVM

Support Vector Machine

XGB

Extreme Gradient Boosting

DSM

Diagnostic and statistical manual of mental disorder

SVC

Support Vector Classifier

mIoU

Mean Intersection Over Union

CARS

Childhood Autism Rating Scale

RFID

Radiofrequency Identification-Enabled

RNN

Recurrent Neural Network

CNN

Convolutional neural network

VR

Virtual reality

EHR

Electronic health record

CBT

Cognitive behavioural therapy

mHealth

Mobile health

NLP

Natural language processing

APD

Auditory processing disorder

LPD

Language processing disorder

ABR

Auditory brainstem response

EDA

Electrodermal activity

HRV

Heart rate variability

IBI

Interbeat interval

LOOCV

Leave-One-Out Cross-Validation

MLP

Multilayer Perceptron

SMM

Stereotypical motor movement

GMFM

Gross Motor Function Measure

rGMFM-66

Reduced Gross Motor Function Measure-66

MDSC

Medical Device Score Calculator

IMU

Inertial measurement unit

nRMSE

Normalised root mean square error

ROM

Range of Motion

GMFCS

Gross Motor Function Classification System

AHA

Assisting Hand Assessment

MACS

Manual Ability Classification System

PDMS-2

Peabody Developmental Motor Scale

MAE

Mean absolute error

LSTM

Long-short-term memory

CAS

Childhood apraxia of speech

DLD

Developmental language disorder

MFCC

Mel-frequency cepstral coefficients-based

AAC

Augmentative and alternative communication

SHAP

SHapley Additive exPlanations

LIME

Local Interpretable Model-Agnostic Explanation

XAI

EXplainable artificial intelligence

LLM

Large language model

CDSS

Clinical decision support system

ADL

Activities of daily living

RCNN

Region-based Convolutional Neural Network

MTCNN

Multi-task Cascaded Convolutional Neural Network

Q`1KNN

K-nearest neighbour

SVM

Support Vector Machine

Author contributions

Conceptualisation: SM, MF and AAW. Material preparation, data collection, analysis and writing of the first draft: SM, AB-J, MF, SR, VS, PANB, SI, AAM, SM and AAW supervision: AAW illustrations: VS, MF. Writing and approval of the final draft of the manuscript: SM, AB-J, MF, SR, VS, PANB, SI, AAM, SM and AAW. All authors: approval of final draft.

Funding

The authors declare that no funds, grants, or other support were received during the preparation of this manuscript.

Availability of data and materials

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

No original data from new patients were collected, consent to participate is not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.