Exploratory Study Uses fNIRS and Machine Learning to Differentiate Autism in Young Children

October 15, 2026 3 Min Read

New Research Explores Brain Imaging and AI to Identify Autism Spectrum Disorder

A recent exploratory study investigated whether functional near-infrared spectroscopy (fNIRS) combined with machine learning algorithms could help distinguish young children with autism spectrum disorder (ASD) from typically developing (TD) peers. The research focused on children aged 2 to 6 years, aiming to assess the potential of this non-invasive brain imaging technique as an adjunct tool for early ASD identification.

Study Design and Methods

The study analyzed brain activation data collected via fNIRS, which measures blood oxygenation changes in the brain, from two different task paradigms. In the first paradigm, 60 children (29 with ASD and 31 TD) participated in three tasks, yielding 24 beta-activation features per participant from eight brain channels. The second paradigm involved 58 children (28 with ASD and 30 TD) completing two tasks, providing 16 beta features per participant.

Researchers applied seven machine learning classifiers—including K-nearest neighbors (KNN), logistic regression, support vector machine, decision tree, random forest, gradient boosting decision trees, and XGBoost—to these features. Data processing steps included normalization, principal component analysis for feature extraction, and leave-one-out cross-validation to evaluate classification performance. Key metrics assessed were accuracy, specificity, sensitivity, and area under the curve (AUC). Permutation testing was also used to determine whether classification results exceeded chance levels.

Key Findings

In the first paradigm, the K-nearest neighbors classifier achieved the highest accuracy at 56.7%, while the decision tree classifier showed the highest AUC of 0.582. The second paradigm produced slightly better numerical results on some metrics, but no classifier demonstrated statistically significant performance above chance according to permutation tests.

Importantly, the study found that the fNIRS-derived features did not provide significant predictive value beyond basic demographic factors such as age and sex. Overall, classification performance was modest, and none of the machine learning models reliably differentiated ASD from TD children in this sample.

Context and Implications

Early diagnosis of ASD is critical for timely intervention, yet current clinical assessments rely heavily on behavioral observations. This study highlights the procedural feasibility of using task-based fNIRS combined with machine learning for exploratory classification purposes. However, the modest results indicate that more research is needed before such approaches can be clinically useful.

Future studies with larger, more balanced cohorts are necessary to validate these findings and to explore whether fNIRS features can contribute meaningful diagnostic information beyond demographic factors. The research underscores the challenges of developing objective, brain-based biomarkers for ASD in young children.

Limitations and Cautions

The study’s exploratory nature and relatively small sample sizes limit the generalizability of the findings. Additionally, the lack of statistically significant classification performance suggests that fNIRS features alone may not be sufficient for reliable ASD identification at this stage. Researchers caution against interpreting these results as evidence of diagnostic capability.

Families and clinicians should continue to rely on established diagnostic criteria and comprehensive clinical evaluations for ASD assessment. Emerging technologies like fNIRS combined with machine learning remain promising but require further development and validation.

For more detailed information, see the original publication: Functional Near-Infrared Spectroscopy Combined with Machine Learning for Exploratory Classification of Autism Spectrum Disorder in Young Children.