Skip to content

Science & research

AI and biomarkers in diagnosis: state of research

Who is it for? For everyone who wants to put reports of “AI diagnosis” or “biomarkers” in autism into context, and for professionals in assessment.

Colour-highlighted program code on a screen
Photo: Markus Spiske markusspiske (CC0 1.0 Universal)

In brief

Reports of AI diagnosis or biomarkers in autism often sound like an imminent solution; the review articles paint a more cautious picture. Programs for facial and behaviour analysis achieved a good but not perfect hit rate, and models for early detection often did worse in independent testing than in development. The evidence does not support replacing professional assessment with standalone AI, and biomarkers from brain measurements are largely still experimental. Autism remains a clinical diagnosis from observation and conversation.

What “biomarker” and “AI” mean here

Autism is diagnosed by behaviour, that is by social communication and repetitive behaviours. Research is therefore looking for measurements from the body, brain or behaviour (biomarkers) that indicate autism earlier or more objectively. It is also looking for computer methods (“artificial intelligence”, machine learning) that recognise patterns in data. [5] To put classical procedures into context, the article Diagnostic procedures: ADOS-2, ADI-R and others informs.

AI on facial and behavioural data

Sun and colleagues screened 40,570 studies. 38 met the criteria, and seven provided enough data for a meta-analysis. The pooled diagnostic odds ratio was 15.9 (wide confidence interval 4.8 to 53.1), and the area under the ROC curve was about 0.86. Procedures using facial features, unstructured play and certain algorithms did better. The authors see AI as a supplement to assessment and call for testing in diverse groups and ethical safeguards. Tests with professionals are missing. [4]

Machine learning for early detection

Onyango and colleagues summarised studies on machine learning for early detection. They covered questionnaires, video, speech, eye tracking, EEG, MRI and sensors. Performance varied greatly. A model based on medical and background information achieved an AUROC of 0.895 in development and 0.790 in independent testing. A practice-oriented AI system achieved a sensitivity of 99.1 % and specificity of 81.6 % for “definite” outputs. The authors stress that the evidence does not support replacing professional assessment. They want to develop AI as support for screening and prioritisation with human oversight. [3] For professionals, a survey of 55 experts is interesting. It found that more experienced colleagues trusted AI screening less. The study is small and preliminary. [2]

EEG, imaging and other brain measurements

De Giacomo and colleagues evaluated 27 studies that combined EEG and fNIRS, that is measurement of brain blood flow, with machine learning. They found recurring patterns but very different measurement set-ups, evaluations and models. Most studies used small datasets, and many models are “black boxes”. This limits transferability and trust. [1]

In a review of biomarkers before behavioural features appear, Wang and colleagues stress that MRI and EEG appear promising for high-risk infants. Many findings are preliminary, expensive and not available everywhere, and genetic biomarkers are complex. The data come mainly from the USA and Europe. [5]

What this means for you

Online tests, apps or services that claim to recognise “autism by AI” do not replace assessment. Experts point to testing in independent groups, oversight by professionals and protection against bias. [3] [4] As a note from the editorial team without a study finding, with such offers you should ask who checks the data, in which groups it was tested and whether a professional puts the result into context. More on questionnaires and tests in the articles Online tests and questionnaires in autism and Neuroscience: what brain research does (not) show.

What we know and what is open

What we know. AI and biomarkers show potential in studies as a supplement. As a replacement for professional assessment they have so far not been demonstrated. [3] [4] [5]

Open.

  • How well the procedures work in independent, diverse and everyday groups. [4] [3]
  • How reliable EEG and imaging findings are in individual children. The datasets are small and inconsistent. [1]
  • Whether the procedures recognise girls, women, adults and people with higher support needs equally well. This has hardly been studied. [4]

You might also be interested in

Sources and literature

  1. De Giacomo A, Palmieri R, Russo EF, Pizzolorusso I, Brandi R, Magistro F, Di Cesare MG, Quattrocelli S, Cardone D, Perpetuini D, Merla A: Machine learning and deep learning applied to EEG and fNIRS for early autism spectrum disorder diagnosis: a systematic review. Frontiers in Psychiatry 2026;17:1668914. https://doi.org/10.3389/fpsyt.2026.1668914 (frei lesbar, CC BY; online 03.02.2026) – Systematic review (27 studies)
  2. Ohmoto Y, Terada K, Iwanaga R, Kumazaki H: Factors Associated With Professionals’ Context-Specific Trust in AI-Based Screening for Autism Spectrum Disorder: Cross-Sectional Survey Study. JMIR Formative Research 2026;10:e98367. https://doi.org/10.2196/98367 (frei lesbar, CC BY; online 25.09.2025) – Cross-sectional survey (55 professionals)
  3. Onyango C, Wanjiku LN, Oketch C: Machine Learning Models for Early Autism Spectrum Disorder Detection: Systematic Review With a Pediatric and Engineering Perspective. Cureus 2026;18(9):e115572. https://doi.org/10.7759/cureus.115572 (frei lesbar, CC BY; online 01.09.2026) – Systematic review
  4. Sun C, McEwan A, Boulton KA, Demetriou EA, Sadozai AK, Lampit A, Guastella AJ: Artificial intelligence for tracking social behaviours and supporting an autism spectrum disorder diagnosis: systematic review and meta-analysis. eBioMedicine 2025;120:105931. https://doi.org/10.1016/j.ebiom.2025.105931 (frei lesbar, CC BY; online 26.09.2025) – Systematic review with meta-analysis (38 studies, 7 in the meta-analysis)
  5. Wang BM, Mills Z, Jones HF, Montgomery JM, Lee KY: Presymptomatic Biological, Structural, and Functional Diagnostic Biomarkers of Autism Spectrum Disorder. Journal of Neurochemistry 2025;169(5):e70088. https://doi.org/10.1111/jnc.70088 (frei lesbar, CC BY-NC; online 01.05.2025) – Systematic review

You might also be interested in

Found a mistake?

If something is factually wrong, out of date or unclear, let us know. You do not have to give any personal details.

Write a noteClose note

What is wrong? A source helps us check it. At least 10 characters.

Only for follow-up questions about this note. Deleted as soon as it has been dealt with. More in the privacy policy.

Support Autismus WissenBooks & further reading