A systematic review of machine learning techniques to address diagnosis and treatment of autism: challenges and opportunities

Rafael Muñoz-Terol; Jesús Peral; Sandra Amador; David Gil

Source-linked AI summary

A systematic review of machine learning techniques to address diagnosis and treatment of autism: challenges and opportunities

Rafael Muñoz-Terol, Jesús Peral, Sandra Amador, David Gil

arXiv:2608.18188v1Aug 18, 2026cs.LGcs.AI

TL;DR

ASD diagnosis and treatment involve complex, variable data, while the effectiveness and trends of machine-learning applications require synthesis. This systematic review of 55 studies from 2017–2023 finds supervised methods dominant and identifies hybrid, multimodal, and multidisciplinary directions for future research.

Problem
Complex ASD-related variables and limited synthesis of recent machine-learning applications motivate reviewing methods used for diagnosis, intervention, and behavioural analysis.
Method
The authors systematically searched major databases for English-language, peer-reviewed ASD machine-learning studies published from 2017 to 2023.
Results
Supervised learning was the most frequently used approach, while deep-learning and hybrid methods emerged as promising directions for ASD applications.
Takeaways & Limitations
Future ASD machine-learning research should pursue multidisciplinary collaboration and customized, context-sensitive approaches for diagnosis, treatment, and intervention.
Takeaways & Limitations
The review notes limited database availability for reproducing experiments and relatively small study samples, with most studies including 0–100 participants.

Abstract

from arXiv · show

Autism spectrum disorder (ASD) is a developmental disability characterized by challenges in social interaction and communication. As the causes of ASD remain unclear, identifying relevant features and hidden correlations is crucial for early diagnosis. This systematic review evaluates 55 studies from 2017 to 2023 on the application of machine learning (ML) techniques to ASD. The primary objective is to examine recent ML applications in ASD research, identifying trends, techniques, and datasets that enhance diagnosis and treatment. Supervised learning methods dominate, as they align well with ASD diagnostic needs; however, the role of deep learning is expanding with greater data availability. Emerging techniques based on hybrid methods, where unsupervised, deep learning, and fuzzy logic could be included, will be interesting to observe in the future. The review highlights key challenges and opportunities, particularly the need for models that can integrate complex data -such as genetic and clinical information- to improve diagnostic accuracy and treatment outcomes. Additionally, incorporating innovative data sources, like wearable devices and biometric sensors, could enable continuous and non-intrusive monitoring, providing a more holistic understanding of ASD. Findings emphasize that addressing current challenges requires interdisciplinary collaboration and expanded datasets tailored to ASD. Future ML models will benefit from broader multimodal data integration, enabling researchers to more comprehensively address the complexities of ASD.

1. Introduction

Machine learning has become a growing approach for ASD detection, diagnosis, observation, and treatment, with supervised, unsupervised, and hybrid methods applied to complex clinical problems. This systematic review addresses the need to evaluate these techniques, their suitability and performance, and ASD dataset availability.

ASD is a developmental disability characterised by impaired social communication and social interactions, alongside restricted and constant patterns of conduct.
ML approaches have demonstrated better results than knowledge-based approaches and are increasingly used for ASD detection and management.
ML algorithms are traditionally classified by learning method into supervised and unsupervised learning, while hybrid methods combine multiple techniques to address challenges single techniques cannot adequately resolve.
Hybrid learning techniques can improve diagnostic accuracy and personalize ASD treatments by combining methods such as neural networks, rule-based systems, and optimization algorithms.Their flexibility and adaptability support classification and analysis of complex data.
The review evaluates ML techniques, their suitability and performance, and the creation and accessibility of datasets for ASD research.

2. Methods and materials

The review systematically searched major databases for ML studies on ASD, using defined research questions and eligibility criteria. Studies were screened through duplicate removal, exclusion procedures, independent review, and PRISMA-guided selection.

Search strategy: Searches of ScienceDirect, Scopus, and MDPI used autism and machine-learning keywords and addressed eight review questions.The keywords included “autism spectrum disorder”, “autism”, “ASD”, and “machine learning”.
Eligibility criteria: Eligible studies were published in English between 2017 and 2023 and addressed ASD-related diagnosis, intervention, or behavioural analysis using ML methods.These criteria were intended to capture recent advancements and technological developments in ML applications to autism research.
Study selection: 163 initial publications were identified, 92 records remained after duplicate removal, and 37 records were excluded before the final selection of 55 studies.The initial publications included 83 from ScienceDirect and 80 from Scopus and MDPI; exclusions included abstracts, conference proceedings, reviews, tutorials, non-English publications, and unrelated records.
Quality assessment: At least two experienced computation and data-science researchers independently reviewed each manuscript’s abstracts, methods, and results for relevance and methodological soundness.The selection process was presented through a PRISMA flowchart outlining all inclusion and exclusion steps.

3. Results … 3.3. RQ3: Which type of ML approach is the most frequently used?

The results organize the review around eight research-question subsections and show that supervised learning is the most frequently used ML approach for ASD detection. The reviewed studies span 2017–2023, with publication and authorship distributions reported by country and supervised methods favored for labeled classification tasks.

3. Results: Eight subsections describe the study conducted to answer the review’s research questions.
3.1. RQ1: Where and when were the previous studies published?: 2017–2023: The reviewed studies were published in high-quality international journals ranked across Journal Citation Report quartiles Q1–Q4.
3.1. RQ1: Where and when were the previous studies published?: 27,27 %: Journals from the Netherlands published the largest number of reviewed research works.
3.2. RQ2: Which countries are the authors of the contributions based in?: Table 3 ranks countries by corresponding-author contributions and remaining-author contributions, based on the institutional countries of those author groups.
3.3. RQ3: Which type of ML approach is the most frequently used?: Supervised learning approaches were used most frequently and achieved the highest percentage in ASD detection tasks because the classification objective was known.
3.3. RQ3: Which type of ML approach is the most frequently used?: Hybrid algorithms are increasing and often complement supervised learning through clustering techniques performed separately.
3.3. RQ3: Which type of ML approach is the most frequently used?: Supervised methods support classification and prediction with labeled data, enabling accurate pattern identification and early ASD detection through markers and symptom classification.
3.3. RQ3: Which type of ML approach is the most frequently used?: “Others” denotes statistical techniques or ensembles of comparative algorithms not included in the review’s other ML categories.

3.4. RQ4: Which ML techniques are the most frequently used? · 3.5. RQ5: Which data types are used most frequently in the selected studies?

Classical ML algorithms were most frequently used in the selected ASD studies, while deep learning may become more prominent as methodologies advance. Brain data predominated among data types, followed by clinical data and eye tracking.

3.4. RQ4: Which ML techniques are the most frequently used?: Classical algorithms were used most frequently across the selected studies.The algorithms were organized across the 2017–2018, 2019–2020, and 2021–2023 intervals.
3.4. RQ4: Which ML techniques are the most frequently used?: Deep learning could become the most frequently used approach in coming years.The passage identifies deep learning as an emerging cutting-edge methodology.
3.4. RQ4: Which ML techniques are the most frequently used?: The “Others” category contains algorithms not assigned to the other categories.This category is defined by exclusion from the remaining algorithm groupings.
3.5. RQ5: Which data types are used most frequently in the selected studies?: Brain data were the most frequently utilized data type, accounting for 45 % of use.The distribution comes from the selected studies’ data-type classification.
3.5. RQ5: Which data types are used most frequently in the selected studies?: Clinical data represented 25 %, followed by eye tracking at 18 %.These were the second- and third-most frequently utilized data types, respectively.
3.5. RQ5: Which data types are used most frequently in the selected studies?: Ballistocardiogram (BCG), prosody, and phenotype were each used only once.The passage contrasts these infrequently used types with the more common brain, clinical, and eye-tracking data.
3.5. RQ5: Which data types are used most frequently in the selected studies?: Three special cases involved studies using different data types within the same study.The passage begins specifying combinations including brain data with eye tracking and clinical data with eye tracking.

3.6. RQ6: How many patients do the selected studies examine? · 3.7. RQ7: Do they normally develop their own databases in their studies? If not, which public databases are the most frequently used?

The reviewed studies most often examined samples of 51–100 or 1001–2000 patients, while database use favored public sources over researcher-developed databases. Public resources included ABIDE, VFDB, Sequence Read Archive, ARIANNA, UCI, KDEF, AGRE, and HCP.

3.6. RQ6: How many patients do the selected studies examine?: 20% of studies examined 51–100 patients, and another 20% examined 1001–2000 patients.Samples of 201–300 patients appeared in 5.45% of studies.
3.6. RQ6: How many patients do the selected studies examine?: Sample sizes of 201–300 patients were reported by 5.45% of studies.This was identified as one of the less frequent sample-size categories.
3.6. RQ6: How many patients do the selected studies examine?: No specific algorithm was found to produce greater bias with the datasets, although this conclusion requires verification using larger data volumes.The analysis correlated ML algorithms with datasets to reduce bias and improve generalisation and extrapolation.
3.7. RQ7: Do they normally develop their own databases in their studies? If not, which public databases are the most frequently used?: 42% of selected manuscripts developed their own databases, whereas 58% used public databases.Studies employed both database strategies.
3.7. RQ7: Do they normally develop their own databases in their studies? If not, which public databases are the most frequently used?: The public databases included Autism Brain Imaging Data Exchange, Virulence Factor Database, Sequence Read Archive, and the ARIANNA database.ARIANNA refers to the Ambiente Di Ricerca Interdisciplinare Per L'Analisi Di Neuroimmagini Nell'Autismo project database.
3.7. RQ7: Do they normally develop their own databases in their studies? If not, which public databases are the most frequently used?: Other public sources included the UCI Machine Learning Repository, KDEF, AGRE, and HCP.These correspond to the University of California Irvine repository, Karolinska Directed Emotional Faces, Autism Genetic Resource Exchange, and Human Connectome Project.

3.8. RQ8: Are ML techniques a good choice for managing ASD?

Across the reviewed studies, ML techniques showed strong performance for ASD research across accuracy, sensitivity, specificity, and AUC, although the best method varied across settings. Particularly high specificity scores, including 100%, and favorable AUC results support ML as a promising approach for ASD research.

Performance assessment: Performance was assessed using the best ML method, accuracy, sensitivity, specificity, and AUC.The review compared five features to analyze the performance of studies applying ML techniques.
Performance assessment: The best ML method varied across study environments because of the wide diversity of methods identified.Both supervised and unsupervised ML methods were represented among the approaches reported as best.
Performance assessment: Specificity scores were high for most ML-based approaches, including scores of 100%, while AUC results also demonstrated good performance.These findings indicate strong performance for applying ML methodology to ASD research.

4. Discussion

The discussion identifies supervised learning as dominant because ASD studies commonly target diagnosis, while emphasizing that model performance depends on choices of data, features, tasks, and algorithms. It highlights data scarcity, bias, generalisability, interpretability, privacy, and interdisciplinary collaboration as central challenges, with larger, more diverse, multimodal, and ethically managed datasets offering opportunities for progress.

ML approaches: ML performance in ASD depends on selecting the appropriate subdomain corpus, features, learning task, and combination of techniques.These variables create challenges because ASD applications involve a large number of computationally relevant factors.
ML approaches: Supervised learning dominates ASD research because many studies focus on diagnosing autism, leading SVM, RF, RL, ANN, and DT to be commonly used.The review also notes that deep learning and hybrid approaches may become more prominent as data availability increases.
Future opportunities: Future research should integrate genetic, clinical, unstructured, and other multimodal data sources to capture information missing from structured records.Free-text clinical notes and DNA biomarkers are identified as promising sources, although genetic analysis remains challenging.
Implementation requirements: Developing effective ASD ML models requires collaboration among computer scientists, clinicians, psychologists, and other stakeholders, alongside strong privacy and informed-consent protections.Sensitive medical data should be collected ethically, with participant or family consent when applicable.
Future opportunities: Federated learning, benchmarking datasets, fine-tuning, and cross-validation could support more robust, general models while protecting patient privacy and representing diverse populations.Benchmarking datasets should include diverse population samples and autism spectrum characteristics.

5. Conclusion

The review finds promising progress in supervised ML for ASD, while hybrid and multimodal approaches offer important future opportunities for diagnosis, treatment, and research. Progress will depend on interpretable models, diverse data integration, and multidisciplinary collaboration.

Conclusion: Future hybrid methods should integrate less structured context-sensitive data, digital twins and synthetic data, and genetic analysis with deep learning.These areas are identified as necessary for improving accessibility and interpretability while uncovering new features and unknown correlations.
Conclusion: Future ASD applications should advance deep learning, natural language processing, semi-supervised learning, reinforcement learning, and interpretable models adapted to clinical datasets.Functional magnetic resonance imaging and eye-tracking data are additional approaches for enhancing diagnostic performance.
Conclusion: Accuracy, F1-score, sensitivity, and specificity consistently indicate that ML and deep learning remain a robust and promising direction for ASD diagnosis.These are identified as the most common evaluation metrics in recent works.
Conclusion: ML applications extend beyond ASD detection and diagnosis to public health, treatment support, and clinical research, requiring collaboration across computer science, psychology, neuroscience, and medicine.Multidisciplinary teams are needed to address ASD complexity and develop customized, context-sensitive applications.

CRediT authorship contribution statement

The authors contributed to writing, investigation, and, for some authors, funding acquisition.

Rafael Muñoz-Terol and Jesús Peral contributed to review and editing, original drafting, investigation, and funding acquisition.
Sandra Amador contributed to review and editing, original drafting, and investigation.
David Gil contributed to review and editing, original drafting, investigation, and funding acquisition.

Ethics declaration

Ethics committee review and informed consent were not required because this literature review used existing published data and involved no direct experimentation on living beings.

Ethics declaration: Ethics committee review and informed consent were not required for this literature review.The study used existing data from published studies and did not involve direct experimentation on living beings.

Funding

The research was funded by the BALLADEER Project and supported by the KOSMOS-UA, BALIDA-AA, and IAEAV projects.

Funding: The research received funding from the BALLADEER Project and support from the KOSMOS-UA, BALIDA-AA, and IAEAV projects.BALLADEER is identified as PROMETEO/2021/088; KOSMOS-UA as PID2024-155363OB-C43; and BALIDA-AA as CIPROM/2024/13.