AI Methods Show Promise in Autism Screening Using Facial Images and DSM-5 Data

October 11, 2026 2 Min Read

Study Explores AI for Autism Spectrum Disorder Screening

Researchers have investigated the potential of artificial intelligence (AI) techniques to assist in screening for autism spectrum disorder (ASD) by analyzing facial images and structured diagnostic checklist data. While AI shows encouraging results in identifying patterns associated with ASD, experts emphasize that these methods are not substitutes for comprehensive clinical evaluations.

Methodology and Data Sources

The study utilized two main data types: facial photographs and DSM-5 checklist records. Specifically, the researchers analyzed 268 DSM-5 records equally divided between children diagnosed with ASD and those without. Additionally, 162 facial images from 92 local children (43 with ASD, 49 without) were included. Publicly available Kaggle image datasets were also used, comprising 2,536 images balanced between ASD and non-ASD classes.

Two AI architectures, EfficientNet-B2 and ResNet50, were trained under two scenarios: one using only the Kaggle dataset and another combining Kaggle images with a small number of local images. The study carefully ensured that images from the same child were not split across training and testing groups to avoid data leakage.

Key Findings

Analysis of the DSM-5 checklist data demonstrated high accuracy, correctly classifying ASD status in 92.5% of cases with an area under the ROC curve (AUC) of 0.984, indicating strong agreement with psychiatrist-assigned diagnoses.

In contrast, AI performance on facial images varied considerably. When trained solely on Kaggle data, ResNet50 achieved an accuracy of 68.4% on local test children after adaptation, while EfficientNet-B2 scored 42.1% without adaptation and improved to 68.4% with adaptation. Using the combined training set, EfficientNet-B2 reached up to 78.6% accuracy post-adaptation, whereas ResNet50 achieved 85.7%. These results, however, are preliminary and based on relatively small local test samples.

Implications for Families and Clinicians

The findings suggest that AI models trained on facial images may capture some features associated with ASD, but their accuracy is influenced by the specific AI architecture, the training dataset, and whether the model is adapted to local populations. The DSM-5 checklist remains a highly reliable tool for screening when used by trained professionals.

Families and clinicians should understand that AI-based screening tools are still in early stages and cannot replace thorough clinical assessments. These technologies might eventually support early identification efforts but require further validation.

Limitations and Future Directions

The study’s authors caution that the image-based AI models have not yet demonstrated clinical validity or the ability to autonomously diagnose ASD. The relatively small size of local test cohorts limits the generalizability of the results. They recommend larger, prospective, multicenter studies with independent reference standards to evaluate AI screening tools before clinical implementation.

Overall, while AI approaches show potential to complement existing ASD screening methods, more rigorous research is needed to confirm their accuracy and utility in diverse populations.

For further details, see the original study published in Health Science Reports: Artificial Intelligence-Based Approaches for Autism Screening Using Facial Images and DSM-5 Checklist Data: A Methodological Study.