Biomarker Discovery for Autism Prediction Using Massive Feature Extraction Based on EEG Signals.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Data Science

Background:

• Autism spectrum disorder (ASD) diagnosis relies on time-consuming behavioral assessments prone to human error.
• Objective biomarkers are crucial for early ASD diagnosis and intervention.
• Electroencephalography (EEG) offers a non-invasive, cost-effective neuroimaging method for ASD research.

Purpose of the Study:

• To develop and validate an objective, feature-based prediction framework for classifying ASD using resting-state EEG.
• To identify key EEG features and channels that discriminate individuals with ASD.
• To enhance the explainability of machine learning models in ASD diagnosis.

Main Methods:

• Utilized the highly comparative time-series analysis (HCTSA) method for extensive feature extraction from resting-state EEG data.
• Implemented a hybrid feature selection approach to identify the most discriminative features.
• Trained and tested machine learning models on a balanced dataset of 56 participants, employing Shapley values for model interpretability.

Main Results:

• The developed framework achieved 100% classification accuracy for ASD with 50 or more selected features.
• Identified specific EEG channels and extracted features that are highly discriminative for ASD.
• Shapley values provided insights into the contribution of different features and channels to the classification outcome.

Conclusions:

• The proposed EEG-based machine learning framework demonstrates high accuracy and potential for objective ASD diagnosis.
• The identified discriminative features and channels offer valuable biomarkers for ASD.
• Further validation on larger, independent cohorts is necessary for clinical translation.