| Popis: |
As a neurodevelopmental disorder with complex pathogenesis, the existing diagnostic methods of autism still only rely on the scale method. In recent years, there have been some methods using machine learning to classify Autism Spectrum Disorders (ASD), and achieved good accuracy. However, the generalization of these models is poor, and the conclusions are inconsistent. The main reason is that most of them use single site dataset or private dataset for analysis, which will lead to one-sided conclusion. They did not analyze the phenotypic features of the dataset, such as handedness, gender, and age. In order to make the obtained brain diagnostic biomarkers of ASD more universal and generalized, instead of analyzing the dataset from a single site, the whole dataset is divided into subgroups according to age and Verbal Intelligence Quotient (VIQ), and then each subgroup is classified and analyzed by Random Forest (RF) model. The experimental results show that if all male subjects are used for classification, the accuracy of classification can only reach about 55%. By using the proposed grouping method and RF model, the classification accuracy for different subgroups will be improved by 3% ~ 17%. Through the analysis of the importance and difference of the features in each subgroup, we can find that the features obtained in the above experiments are closely related to the functions of speech, emotion, auditory and visual information processing. This may partly explain why ASD patients have speech, social disorder, repetitive behavior and narrow interests. The classification methods proposed and diagnostic biomarkers obtained in this paper may provide some reference for the clinical diagnosis and early treatment of ASD. |
