Deep Learning-Based Early Autism Spectrum Disorder Detection and Stage Classification Using SHAP Explainable AI Framework

Abstract

The Early Autism Spectrum Disorder (ASD) detection based on Deep Learning is critical to allow timely clinical intervention and personalized therapy planning. The proposed study is an explainable deep learning framework to detect early ASD and classify the stage using SHAP (SHapley Additive exPlanations) to improve model transparency and clinical trust. The main aim is to come up with a strong and interpretable model that can effectively determine the presence of ASD and the stages of its severity using heterogeneous data. The methodology is a hybrid deep learning architecture that combines Convolutional Neural Networks (CNN) to extract features and Long Short-Term Memory (LSTM) networks to analyze temporal patterns. The model is trained using real-world behavioral and clinical data, and includes preprocessing, feature normalization, and class balancing methods. SHAP is used to explain the contribution of features and to identify important behavioral indicators that affect predictions. Experimental results demonstrate high classification accuracy, improved precision and recall, and reliable stage-wise discrimination compared to conventional machine learning models. The results indicate that explainable AI can greatly increase the interpretability of the model without affecting performance. The suggested framework will equip clinicians with practical information about the decision-making process, which will be used to support early diagnosis and intervention plans. To conclude, this paper presents a scalable and transparent AI-driven solution to ASD detection, which bridges the gap between high-performance deep learning models and the interpretability requirement in healthcare applications. 