Functional near-infrared spectroscopy-based machine learning techniques for autism spectrum disorder diagnosis: a
Xiawei Li1, Fuli Yan1, Yinghua Wu2
1College of Rehabilitation Medicine, Jiamusi University, Jiamusi 154007, Heilongjiang, China; Children Rehabilitation Nerve Laboratory, Jiamusi University ,Jiamusi 154007, Heilongjiang, China; The Third Affiliated Hospital of Jiamusi University, Jiamusi 154007, Heilongjiang, China.
Background:
Diagnosis of autism spectrum disorder (ASD) primarily relies on subjective behavioral assessments, posing risks of delayed and missed diagnoses. The combination of functional near-infrared spectroscopy (fNIRS) technology and machine learning (ML) offers a novel approach for objective ASD identification; however, its overall diagnostic efficacy and key influencing factors require systematic evaluation.
Methods:
Following the PRISMA guidelines, we systematically searched PubMed, EMBASE, Web of Science, The Cochrane Library, and Wiley Online Library databases to comprehensively collect English-language literature on fNIRS-based machine learning techniques for ASD diagnosis published from the inception of our database to December 2025. Data were extracted to calculate pooled sensitivity (sen), specificity (spe), positive likelihood ratio (LR + ), negative likelihood ratio (LR-), diagnostic odds ratio (DOR), and their 95% confidence intervals (CI). Pooled receiver operating characteristic (ROC) curves were plotted, and area under the curve (AUC) was calculated to assess diagnostic value. The I2 test assessed study heterogeneity, which was explored through meta-regression and subgroup analyses. Publication bias was evaluated using Deeks funnel plot asymmetry tests. The review protocol was registered in PROSPERO (CRD420251250866).
Results:
A total of 17 studies were included in the systematic review, with 15 studies included in the meta-analysis. The pooled analysis showed that the sensitivity of fNIRS-based ML techniques for diagnosing ASD was 0.92 (95% CI 0.87-0.95),with a specificity of 0.94 (95% CI 0.90-0.97) and a composite area under the receiver operating characteristic curve (AUC) of 0.98 (95% CI 0.96-0.99). Based on these findings, fNIRS-based ML techniques demonstrate good diagnostic value for ASD.
Conclusion:
fNIRS-based ML techniques demonstrate excellent diagnostic accuracy for ASD, presenting a highly promising objective diagnostic tool. Future research should focus on establishing standardized, multicenter datasets and analytical workflows, while exploring multimodal data fusion and model interpretability to advance this technology toward stable, reliable clinical diagnostic support.
