An integrated data-driven analysis-based deep learning framework for early autism detection in children to improve diagnostic performance

Autism spectrum disorder (ASD) children must be recognised early to obtain prompt care, promote development, and reduce long-term issues. This research provides a VGG16 and ResNet50-based data-driven deep learning system for early ASD screening using facial picture data. The study meticulously normalises, augments, and selects features using chi-square methods to ensure high-quality inputs and low dataset variability. Hyperparameter adjustment optimises model performance and five-fold cross-validation provides robust evaluation. VGG16 can recognise complex face characteristics with 87% accuracy for autistic classifications due to its precision and recall measures. Bio-inspired optimisation improves classification, helping ResNet50 outperform training epochs. Despite these advances, multimodal inputs are still needed for complete analysis due to the limits of facial data and the diversity of datasets. Deep learning models with feature selection can improve diagnostic precision, reduce false positives, and enable clinical real-time ASD screening. The proposed framework speeds diagnosis and is adaptable to varied healthcare circumstances. Future studies will focus on behavioural and genetic data, expandable artificial intelligence (XAI) for interpretability, and larger datasets for robustness. A scalable and effective ASD diagnosis using AI shows the transformative potential of AI in healthcare.