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

Area of Science:

• Neuroscience
• Computer Science
• Machine Learning

Background:

• Autism Spectrum Disorder (ASD) detection remains a challenge, particularly with limited or imbalanced datasets.
• Traditional machine learning and deep learning models often struggle with generalizability across different data sources.
• Developing robust and accurate ASD screening tools is crucial for early intervention.

Purpose of the Study:

• To present a novel transfer learning approach for Autism Spectrum Disorder (ASD) detection.
• To evaluate the performance of Deep Neural Networks (DNN) in cross-dataset ASD classification.
• To identify common behavioral indicators for ASD using explainable AI.

Main Methods:

• A baseline model was trained on a Saudi Arabian toddler ASD screening dataset, employing data augmentation (SMOTE) and a DNN architecture with regularization and dropout.
• The trained DNN model's knowledge was transferred to two independent ASD datasets.
• Model performance was assessed using standard metrics and explainable AI techniques.

Main Results:

• The DNN architecture significantly outperformed other models like LSTM and Attention LSTM.
• Transfer learning demonstrated enhanced performance, especially when dealing with limited training data.
• Explainable AI provided insights into key ASD classification features across different populations.

Conclusions:

• Transfer learning is an effective strategy for cross-dataset ASD classification.
• Common behavioral indicators for ASD appear to exist across diverse demographic and data collection contexts.
• This approach holds promise for improving the accuracy and generalizability of ASD screening tools.