Abstract
The accurate diagnosis of autism spectrum disorder (ASD) is of great practical significance in clinical practice. The spontaneous hemodynamic fluctuations were collected by functional near-infrared spectroscopy (fNIRS) from the bilateral frontal and temporal cortices of typically developing (TD) children and children with ASD. Since traditional machine learning and deep learning methods cannot make full use of the potential spatial dependence between variable pairs, and require a long time series to diagnose ASD. Therefore, we use adaptive spatiotemporal graph convolution network (ASGCN) and short time series to classify ASD and TD. To capture spatial and temporal features of fNIRS multivariable time series without the pre-defined graph, we combined the improved adaptive graph convolution network (GCN) and gated recurrent units (GRU). We conducted a series of experiments on the fNIRS dataset, and found that only using 2.1 s short time series could achieve high precision classification, with an accuracy of 95.4%. This suggests that our approach may have the potential to detect pathological signals in autism patients within 2.1 s. In different brain regions, the left frontal lobe has the best classification effect, and the abnormalities occur more frequently in left hemisphere and frontal lobe region. Moreover, we also found that there were correlations between multiple channels, which had different degrees of influence on the classification of ASD. From this, we can also generalize to a wider range, there may be potential correlations between different brain regions. This may help to better understand the cortical mechanism of ASD.
Introduction
Early and accurate diagnosis of Autism Spectrum Disorder (ASD) is critical for effective intervention, yet traditional clinical methods can be subjective and time-consuming. Functional near-infrared spectroscopy (fNIRS) has emerged as a promising non-invasive tool because it is portable and relatively insensitive to motion—making it ideal for children.
However, existing machine learning models for fNIRS data face two major hurdles: standard deep learning models often fail to capture latent spatial dependencies between different brain regions, and most models require long time series (often minutes of data) to reach high accuracy. This study introduces an Adaptive Spatiotemporal Graph Convolution Network (ASGCN) designed to learn the brain's connectivity graph autonomously and perform accurate classification using only a few seconds of data.
Materials and Methods
Data Acquisition and Participants
The study utilized a dataset of spontaneous hemodynamic fluctuations from 47 children: 25 with ASD (average age 9.3) and 22 typically developing (TD) children (average age 9.5). Data was collected from 44 channels covering the bilateral frontal and temporal cortices, focusing on oxygenated hemoglobin ($HbO_2$) and deoxygenated hemoglobin ($Hb$).
Data Preprocessing
To handle the relatively small number of participants and the goal of short-term diagnosis, the researchers employed a sliding window technique. This process broke down 8-minute recordings into thousands of overlapping short-time segments (1.4s, 2.1s, and 3.5s), effectively augmenting the dataset for deep learning training.
ASGCN Architecture
The core innovation is the ASGCN, which integrates:
- Adaptive Graph Learning (AGL): Automatically learns the connections (edges) between fNIRS channels (nodes) during training, bypassing the need for pre-defined connectivity maps.
- Node Weight Adaptive Learning (NWAL): Assigns specific weights to each brain region to capture unique spatial dependencies.
- Temporal Capture: Combines the GCN with Gated Recurrent Units (GRU) to process the time-series aspect, allowing the model to understand how brain signals evolve over short intervals.
Results
Classification Performance
The model achieved a peak accuracy of 95.4% using a 2.1-second sliding window and $HbO_2$ data. This suggests that pathological signals of ASD are embedded in very brief, spontaneous fluctuations of brain activity. The ASGCN significantly outperformed baseline models, including standard GRUs (65% AUC) and other advanced architectures like Graph Wavenet (88% AUC).
Regional and Biomarker Sensitivity
- Left Frontal Lobe Dominance: Testing different brain regions revealed that the left frontal lobe yielded the highest accuracy (95.9%), reinforcing its known role in language processing and social cognition abnormalities in ASD.
- $HbO_2$ vs. $Hb$: The model performed significantly better using oxygenated hemoglobin ($HbO_2$) data, identifying it as a more sensitive biomarker for ASD classification than deoxygenated hemoglobin.
Discussion and Conclusion
The ASGCN model demonstrates that high-accuracy ASD identification is possible with remarkably short data bursts (2.1s), reducing the required data length by 40% compared to previous state-of-the-art studies. While the sample size of 47 subjects is a limitation, the model's ability to automatically discover functional connectivity patterns—such as the high correlation between specific channel pairs—offers a "data-first" approach to understanding the cortical mechanisms of ASD. Future work will focus on validating these findings in larger, more diverse cohorts and applying the ASGCN architecture to other multivariate time series tasks.
Summary updated