Identification of autism spectrum disorder based on short-term spontaneous hemodynamic fluctuations using deep learning in a multi-layer neural network.

OBJECTIVE

To classify children with autism spectrum disorder (ASD) and typical development (TD) using short-term spontaneous hemodynamic fluctuations and to explore the abnormality of inferior frontal gyrus and temporal lobe in ASD.

METHODS

25 ASD children and 22 TD children were measured with functional near-infrared spectroscopy located on the inferior frontal gyrus and temporal lobe. To extract features used to classify ASD and TD, a multi-layer neural network was applied, combining with a three-layer convolutional neural network, a layer of long and short-term memory network (LSTM) and a layer of LSTM with Attention mechanism. In order to shorten the time of data collection and get more information from limited samples, a sliding window with 3.5 s width was utilized after comparisons, and numerous short (3.5 s) fNIRS time series were then obtained and used as the input of the multi-layer neural network.

RESULTS

A good classification between ASD and TD was obtained with considerably high accuracy by using a multi-layer neural network in different brain regions, especially in the left temporal lobe, where sensitivity of 90.6% and specificity of 97.5% achieved.

CONCLUSIONS

The "CLAttention" multi-layer neural network has the potential to excavate more meaningful features to distinguish between ASD and TD. Moreover, the temporal lobe may be worth further study.

SIGNIFICANCE: The findings in this study may have implications for rapid diagnosis of children with ASD and provide a new perspective for future medical diagnosis.