A novel hybrid deep learning model using MEResNext for autism spectrum disorder detection

Area of Science:

• Neurology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Autism spectrum disorder (ASD) is a neurological condition characterized by challenges in social interaction, communication, and repetitive behaviors.
• Early identification and intervention are crucial for managing ASD, potentially reducing the need for extensive evaluations and costly medical procedures.
• Timely diagnosis of ASD can mitigate long-term effects and reduce symptom severity.

Purpose of the Study:

• To develop and evaluate a hybrid deep-learning method for accurate autism spectrum disorder (ASD) detection.
• To improve upon existing models for ASD diagnosis through advanced feature selection and classification techniques.

Main Methods:

• A hybrid deep-learning approach involving three stages: pre-processing, feature selection, and ASD detection.
• Data pre-processing utilized Yeo-Jhonson transformation for noise and artifact removal.
• Feature selection was performed using a Double Exponential Smoothing-Elk Herd Optimizer (DeSEHO), integrating Double Exponential Smoothing (DES) with the Elk Herd Optimizer (EHO).
• ASD detection was conducted using Moments Embedding ResNeXt (MEResNeXt), a fusion of Moments Embedding Network (MoNet) and ResNeXt.

Main Results:

• The proposed MEResNeXt model demonstrated superior performance compared to traditional models.
• Achieved an accuracy of 95.3%, sensitivity of 96.5%, and specificity of 94.8% in ASD detection.
• The hybrid deep-learning method effectively identified key features for ASD classification.

Conclusions:

• The MEResNeXt model represents a significant advancement in deep learning for autism spectrum disorder detection.
• The study highlights the potential of hybrid AI approaches for precise and efficient neurological disorder diagnosis.
• Early and accurate ASD detection using advanced computational methods can lead to better patient outcomes.