Introduction

1. ASD (Autism Spectrum Disorder) affects:
Social communication and behavior (e.g., repetitive patterns).

Diagnosis today is clinical, based on observation, interviews, and tests that generally:
a. Can take a long time;
b. Depend heavily on human interpretation;
c. May vary from one professional to another.

2. Where do EEG and the computer come in? EEG measures the brain’s electrical activity through electrodes on the scalp.

The idea is to use computers + machine learning to:
a. Analyze these signals;
b. Detect patterns associated with ASD.

This is called CAD (Computer-Aided Diagnosis).

Problem:
AI models are often “black boxes”: they give an output, but do not explain why.

Studies using EEG in ASD diagnosis have existed since 2014:

2014: Early studies focused on robust feature extraction and ASD classification from EEG were already being published. Robust features for the automatic identification of autism spectrum disorder in children

2015: Research on EEG biomarker validation gained strength. Robust disruptions in electroencephalogram cortical oscillations and large-scale functional networks in autism

2020–2023: Multiple works used machine learning and advanced techniques to improve ASD detection and diagnosis. A robust method for early diagnosis of autism spectrum disorder from EEG signals based on feature selection and DBSCAN method

2024–2025: Deep network models and neural variability analyses continue to expand the field. Harnessing Trial-to-Trial Variability of EEG Spectral Characteristics to Understand Autism

SUMMARY OF THESE ARTICLES AND STUDIES:

EEG shows clear differences in brain rhythms (e.g., alpha, delta) and functional connectivity between individuals with ASD and controls.

Research has an increasing focus on explainable and reproducible biomarkers, not only classifiers.

Recent studies also explore early detection in infants (before clinical symptoms).

Results

3. What does this 2025 study propose that is new?

They created a model called GASTCN: Graph Attention Spectral-Temporal Convolutional Network which is a deep neural network that analyzes EEG across time, frequency, and relationships between electrodes, while also explaining its decisions.

4. What does the model actually analyze in EEG?

a. Time and frequency at the same time: The EEG signal is transformed into spectrograms.

This shows:

(They use the Short-Time Fourier Transform (STFT) for this)
Think of it as a “color map” of the brain over time.

b. DSE block (Diversified Selection Extraction) This block does two things:

It expands information:

In other words: It helps the model not lose important details, while also not being confused by noise.

c. Relationships between electrodes (graph attention):
The brain works as a network, not as isolated points, and the model learns which electrodes “communicate” with each other:

They use "attention guided by a 'learnable token': the model itself learns where to focus 
on the scalp and which EEG channels are most important.

5. Does the model work well? Yes, very well.

They tested it using 5-fold stratified cross-validation, which means: The model was trained and tested multiple times on different splits of the data, avoiding “luck” or bias.

Results:

6. And the most important part: does it explain the result?
Yes. This is the key advantage.

They used two explainability techniques:

With this, they discovered:

Main ASD biomarkers:

Delta and alpha bands at electrode C3 (left central brain region).
That is: it is not just “the AI said it is ASD”, but “these frequencies, at this brain location, were decisive”.

BUT ARE THERE PREVIOUS STUDIES ADDRESSING AUTISM DIAGNOSIS WITH EEG?

The answer is: YES

HOW IS THIS DIAGNOSIS POSSIBLE USING EEG?

First, I will explain how EEG works and how ASD can be identified in such an exam:

1. Atypical organization of baseline rhythms

In neurotypical adults, one expects: A well-organized posterior dominant alpha rhythm, symmetric and with good reactivity (Meaning: open the eyes: alpha decreases)

In many autistic individuals (including adults):

IN OTHER WORDS: “This brain did not organize its resting rhythm according to the standard pattern.”

2. Excess focal or diffuse slow activity (without epilepsy)
Very common in ASD: Delta/theta in central or frontal regions (especially without epileptiform discharges) appears at rest, not only during drowsiness.

This is striking because:

Today we know this relates to atypical cortical development and altered connectivity

This is not classical pathology, but it is neurodevelopmentally atypical.

3. Persistent functional asymmetries Another classic point:
Clear, non-transient interhemispheric differences not explained by lesion, especially in central (C3/C4), temporal and frontal regions.
This indicates atypical connectivity and lateralization.

4. Abnormal reactivity to simple stimulation
During EEG:
a. Eye opening/closing;
b. Hyperventilation;
c. Intermittent photic stimulation.

In many autistic individuals
a. Exaggerated or reduced responses;
b. Unusual latency;
c. “Strange” desynchronization.

This suggests different sensory processing, which is now recognized as central in ASD.

5. Absence of the expected “normal pathological” pattern:
This “in-between” state is typical of ASD.

Figure 1 Multimodal EEG–AI pipeline for ASD detection. Raw EEG signals are transformed into time–frequency representations (STFT), processed by feature selection (DSE) and graph attention to capture brain network structure, analyzed by the GASTCN model, and interpreted via explainable AI (SHAP) to extract neurobiological biomarkers.

EEG and Artificial Intelligence for Autism Diagnosis
From clinical neurophysiology to explainable deep learning (2014–2025)
Block What it is What EEG shows How AI analyzes it Relevance for ASD
ASD (clinical basis) Neurodevelopmental disorder affecting social communication and behavior Atypically organized brain rhythms — Clinical diagnosis is subjective and slow
EEG in ASD Recording of brain electrical activity via scalp electrodes Alpha, delta, theta, beta bands and functional connectivity Signal preprocessing and feature extraction Reveals real brain function
Classical biomarkers Stable neurophysiological patterns associated with ASD Altered alpha, excess delta/theta, C3–C4 asymmetries, abnormal reactivity Relevant feature selection Reflect atypical cortical development and connectivity
AI + EEG (CAD) Computer-aided diagnosis using machine learning Raw and spectral EEG signals Classifiers and neural networks Reduces bias and increases reproducibility
2025 Model (GASTCN) Spatial-temporal deep neural network with graph attention Time-frequency maps + relationships between electrodes STFT + DSE + Graph Attention Detects true spatiotemporal ASD patterns
Explainability Methods that explain why the model made a decision Which frequency bands and channels mattered most SHAP + weighted spectral power Identifies neurobiological biomarkers
Clinical performance How well the model distinguishes ASD from controls Clear separation between ASD and non-ASD Stratified cross-validation AUROC 94.3%, Sensitivity 96.1%
Based on EEG studies in ASD (2014–2025) and the GASTCN model (2025)

Table 1 Integration between EEG, neurophysiological biomarkers and explainable deep learning models for Autism Spectrum Disorder (ASD) diagnosis. The table summarizes the evolution from classical EEG findings to modern graph-based deep neural networks (GASTCN) with interpretable biomarkers.