Harnessing Explainable AI for Next-Generation Autism Diagnosis and Severity Assessment

Autism Spectrum Disorder (ASD) diagnosis and intervention remain critical challenges due to the complex interplay of behavioural traits and the lack of definitive diagnostic tools.This study introduces an innovative Multilayer Perceptron (MLP) framework for ASD prediction and severity classification across diverse age groups-toddlers, children, adolescents, and adults-using non-clinical datasets from the Kaggle and UCI repositories.The dual-stage model integrates ASD classification and severity-level categorization (mild, moderate, and severe) while leveraging Explainable AI techniques, specifically SHapley Additive extensions (SHAP), to provide insights into feature importance.The model demonstrated exceptional performance, achieving an F1-score exceeding 0.97 in toddler and adult classifications and maintaining robust accuracy across all severity levels.Severity classification showed similar efficacy, with precision and recall exceeding 98% for toddlers and 96% for adults, significantly.The outperforming existing models include the stacked sparse autoencoder with a softmax classifier and salience network-based methods.SHAP analysis highlights key behavioural and demographic features across age groups and severity categories, underscoring the framework's potential to tailor interventions.This lightweight, scalable model bridges critical gaps in age-inclusive and nonclinical ASD diagnostics, offering a transformative solution for early detection and personalized care.Future work will refine the model's adaptability to real-world scenarios and explore cross-cultural applications.