.png){kind=icon}
.png){kind=icon}
Return
A unified frequency–temporal–spatial EEG framework for Autism Spectrum Disorder Diagnosis
.png){kind=icon}
PRE
AI
Z
S
G
李
DOI：10.1016/j.bspc.2026.111195.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Objective detection of Autism Spectrum Disorder (ASD) using Electroencephalography (EEG) remains a challenge due to the non-stationary nature of neural signals and the complex interplay of spectral–temporal–spatial biomarkers. Conventional methods typically analyze these domains in isolation, potentially missing critical cross-domain interactions associated with ASD pathology. In this paper, we introduce a unified deep learning framework for accurate ASD classification, termed FTSNet, which jointly models frequency, temporal, and spatial neural dynamics. The proposed architecture explicitly models multi-dimensional neural dynamics through three integrated components: (1) a frequency-aware attention mechanism to adaptively weight disorder-relevant spectral bands; (2) a multi-scale temporal convolution module to capture transient features across varying time windows; and (3) a graph-based module to extract topological brain connectivity patterns. Extensive experiments were conducted on a collected clinical EEG dataset involving individuals with ASD and typically developing controls. FTSNet achieved a classification accuracy of 94.8%, demonstrating superior performance compared to existing machine learning and deep learning baselines. Furthermore, saliency analysis revealed that FTSNet captures neurophysiologically interpretable features, including specific altered functional connectivity patterns. These findings suggest that FTSNet offers a robust and effective approach for automated EEG-based ASD screening, holding significant promise for computer-aided clinical diagnosis systems.
Journal
IF:
4.9
Papers:
9.7K
Citations:
2.4W
Organization
Cited Papers