# Explainable AI uncovers novel EEG microstate candidate neurophysiological markers for autism spectrum disorder

## TL;DR

This study uses EEG microstate analysis and explainable AI to identify potential brain activity patterns that could help diagnose autism spectrum disorder.

## Contribution

The novel contribution is the use of multidomain microstate-informed features with interpretable AI to identify robust biomarkers for ASD.

## Key findings

- XGBoost achieved 80.87% accuracy using multidomain microstate features for ASD classification.
- SHAP analysis identified 20 discriminative features, including spectral and temporal metrics, as potential biomarkers.
- Retraining on top features confirmed their robustness with 80.34% accuracy and statistical validation.

## Abstract

Autism spectrum disorder (ASD) is a complex neurodevelopmental condition characterized by atypical brain connectivity and impaired cognitive flexibility. Electroencephalography (EEG) based microstate analysis provides insight into the rapid temporal dynamics of brain networks, offering potential biomarkers for ASD.

This study proposes an interpretable classification framework for ASD diagnosis using multidomain microstate-informed features derived from EEG, integrating temporal, spectral, complexity-based, and higher-order metrics to comprehensively characterize brain dynamics.

Resting state EEG data from 56 participants (28 with ASD and 28 neurotypical controls; age range: 18–68 years) from the publicly available Sheffield dataset were preprocessed and segmented into microstates using a data-driven clustering approach. From these microstate sequences, we extracted a rich set of features across four domains: (i) temporal, (ii) spectral, (iii) temporal complexity, and (iv) higher-order metrics. Multiple classifiers were evaluated using 10-fold cross-validation, with hyperparameter tuning via a randomized search.

Among all classifiers, XGBoost achieved the highest performance, with an accuracy of 80.87% when utilizing the complete multidomain feature set, significantly outperforming single domain models. Explainable AI analysis using SHapley Additive exPlanations (SHAP) identified the top 20 discriminative features, including fractional occupancy derivative for microstate 3, delta-band power in states 1 and 3, and mean inter-transition interval. Retraining XGBoost on these SHAP-selected features yielded 80.34% accuracy, confirming their robustness as potential biomarkers. Statistical validation via Mann–Whitney U-tests and effect size measures further established their significance.

The findings from the study demonstrated that microstate-informed features capturing temporal instability, transition unpredictability, and spectral alterations serve as clinically relevant and interpretable candidate neurophysiological markers of ASD, offering translational potential for objective diagnosis, treatment monitoring, and personalized interventions.