Deep transfer learning and explainable AI framework for autism spectrum disorder detection across multiple datasets

Introduction: This paper presents a transfer learning approach for Autism Spectrum Disorder (ASD) detection using Deep Neural Networks (DNN) across three distinct datasets. Methods: A baseline was established by training multiple machine learning and deep learning models on a toddler ASD screening dataset from Saudi Arabia, augmented with the Synthetic Minority Oversampling Technique (SMOTE) to address class imbalance. The DNN architecture featured regularization and dropout layers. The trained model was then leveraged by transferring learned knowledge to two additional ASD datasets. Model performance was analyzed through standard metrics and explainable AI techniques. Results: The DNN architecture outperformed other models (i.e., LSTM and Attention LSTM). Transfer learning demonstrated improved performance with limited training data. Explainable AI techniques provided insights into key features for ASD classification across different populations. Discussion: Results indicate the efficacy of transfer learning for cross-dataset ASD classification, suggesting the presence of common behavioral indicators despite demographic and data collection differences.