A hybrid model for the classification of Autism Spectrum Disorder using Mu rhythm in EEG

Area of Science:

• Neuroscience
• Computational Biology
• Biomedical Engineering

Background:

• Autism Spectrum Disorder (ASD) diagnosis relies on subjective evaluations, lacking objectivity.
• Machine learning (ML) and deep learning (DL) offer potential for objective ASD classification.
• Electroencephalography (EEG) data can provide objective biomarkers for ASD.

Purpose of the Study:

• To improve classification accuracy between individuals with ASD and typically developing (TD) individuals using minimal EEG channels.
• To explore the efficacy of ML and DL models in analyzing EEG data for ASD detection.
• To investigate the impact of non-linear features and image transformation techniques on classification performance.

Main Methods:

• Extracted non-linear features from time and frequency domains of EEG data.
• Transformed 1D EEG data into images using Independent Component Analysis-Second Order Blind Identification (ICA-SOBI), Spectrogram, and Continuous Wavelet Transform (CWT).
• Applied ML models (e.g., Stacking Classifier) and DL models, including a hybrid approach combining DL and ML.

Main Results:

• ML models with non-linear features achieved 78% accuracy; incorporating entropy features improved it to 81.4%.
• DL models using SOBI, CWT, and spectrogram plots achieved 75% accuracy.
• A hybrid model combining DL features (spectrogram, CWT) with ML achieved 94% accuracy, further improved to 96.9% with entropy and fuzzy entropy features.

Conclusions:

• ML and DL techniques significantly enhance ASD and TD classification accuracy.
• Utilizing a minimal set of EEG channels is effective for ASD classification.
• The hybrid model demonstrates superior performance, highlighting its potential for objective ASD diagnosis.