A dual-modal approach for detecting and classifying autism spectrum disorder using behavioral and facial features

Background: Autism spectrum disorder (ASD) is a neurodevelopmental condition that significantly affects social connection, behavior, and knowledge acquisition. Despite increasing global prevalence, timely diagnosis remains challenging due to heterogeneity in clinical presentation. Aim: The aim of this study is to develop a dual-modal framework for early detection of ASD by analyzing behavioral assessment and image data. Methods: The proposed framework consists of two independent yet complementary modules. In the behavioral module, questionnaire responses and assessment data were analyzed using an artificial neural network classifier to predict the likelihood of ASD. In the visual module, facial images were analyzed using a DenseNet121-based deep learning model with transfer learning to detect ASD-related traits. Each module independently estimates ASD probability and categorizes severity levels. Results: The DenseNet121 model achieved strong performance in image-based ASD detection, with 91.16% (95% confidence interval [CI]: 87.8–94.2) accuracy, 91.2% (95% CI: 87.8–94.2) sensitivity, and 89.8% specificity, including when trained on a relatively small dataset. Independent training of the two modules may improve robustness and reduce modality-specific bias. Conclusion: The proposed framework demonstrates potential for enhancing early ASD detection using dual modalities. The findings support the use of deep learning-based approaches to improve detection accuracy. Relevance for patients: Early screening of ASD can facilitate timely interventions and personalized care strategies. This method offers a noninvasive, data-driven approach that may support caregivers and healthcare systems in informed decision-making, ultimately benefiting individuals with ASD and their families.