AB-ALSTM Model for Autism Early Warning Based on Behavioral Characteristics
ShanShan Zhong, Jiaye Lin, Dayu Huang, Manning Lu, Daifeng Li · 2021
According to the Centers for Disease Control in China, one in 160 children may have autism spectrum disorder (ASD). The autistic community has been widely concerned by the society. Social networking has become an essential channel for autistic communities to express their voices and share their experiences, but few studies have focused on the influence of this information explosion, which is generated by autistic groups, towards autism researches. Therefore, in this paper, we propose an AB-ALSTM model for analyzing autism behavior data from Baidu Autism Post Bar (BAPB), which is one of the largest online autism community in China. The main idea is to extract key influence factors of autism related behavioral characteristics from large-scale unstructured text of BAPB. In AB-ALSTM, ALSTM (Long-short term memory with attention mechanism) is firstly proposed to extract meaningful behavior patterns related with autism symptoms, then AdaBoost (AB) is ensembled to optimize and refine the extracted patterns. Extensive experiments are conducted and experimental results show that the proposed model has an accuracy of 0.89 in identifying key influence factors of autism symptoms, which significantly outperformances the other state-of-the-art baseline models. The research can be used to design questionnaires to calculate the behavioral similarity between suspected autism and autism characteristics, allowing parents who lack knowledge of specialized fields such as psychology and medicine to accurately grasp their children's symptoms and realize early autism warning. Furthermore, the proposed model can be further optimized for the initial medical diagnosis of suspected autism, as well as to achieve accurate matching between doctors and patients, intelligent information recommendation and so on.