Autism spectrum disorder (ASD) is a heterogeneous neurodevelopmental disorder that requires early diagnosis for better intervention. However, current clinical behavioural examinations are time-consuming and prone to human error. Objective and effective biomarkers are essential for the diagnosis and prognosis of the disorder. Electroencephalography (EEG) is a non-invasive and inexpensive brain-imaging technique that is widely applied in the diagnosis of ASD. Feature-based methods have shown better performance in EEG-based applications. Here, we present a prediction framework based on massive feature extraction using the highly comparative time-series analysis (HCTSA) method and a hybrid feature selection method for the classification of ASD from resting-state EEG. Machine-learning models are trained and tested on a different number of selected features. Our models demonstrated 100% accuracy with ≥50 features on a balanced dataset of 56 participants. The most discriminating EEG channels and features were used in the prediction process, as well as those using Shapley values to provide explainability of our framework. Whilst these results are promising, we acknowledge the limitations of a single small-scale dataset and emphasise the need for validation on larger independent cohorts before clinical translation.
Hafeez et al. (Mon,) studied this question.