Study on the method of using attention mechanism and meta-learning to diagnose autism under small sample multi-omics condition
10.12173/j.issn.1005-0698.202412113
- VernacularTitle:小样本多组学数据下利用注意力和元学习进行孤独症诊断的方法研究
- Author:
Qi WANG
1
;
Kun XIE
;
Xuezhi LIANG
;
Xiangyang LUO
;
Ying LIU
;
Wen CHEN
Author Information
1. 中山大学公共卫生学院医学统计学系(广州 510080)
- Publication Type:Journal Article
- Keywords:
Autism diagnosis;
Attention mechanism;
Meta learning;
Small sample;
Multi-omics data integration
- From:
Chinese Journal of Pharmacoepidemiology
2025;34(8):887-896
- CountryChina
- Language:Chinese
-
Abstract:
Objective To develop a deep learning method for small sample multi-omics data using attention mechanism and Meta-learning for the establishment of autism diagnosis model.Methods MLAN(Meta-learning based attentive network)consisting of the omics feature pre-reduction module,the multi-omics data fusion and feature learning module,and the parameter optimization module was designed.Firstly,differential expression analysis was performed on high-dimensional multi-omics data to preliminarily screen out unimportant features.Secondly,a multi-channel attention mechanism was used to learn the importances of every set of omics data and to realize data fusion,and a two-layer fully connected network was constructed to further extract latent features and realize the diagnosis task.Finally,the Meta-learning algorithm Reptile was used to optimize the initial parameters of the above model to obtain the optimal parameters.A total of 58 children's saliva samples were collected,including 21 children diagnosed with autism,12 children with social disorders,and 25 healthy controls,and the protein and metabolomics data were detected by mass spectrometry.All data were randomly divided into training set and test set by 4∶1,and the training set was divided into training data and validation data in the same way for model training and validation.The test set was used for the final evaluation of the model effect.Five baseline models and three ablated models were constructed and evaluated along with MLAN based on metrics including multi-classification accuracy,F1-macro and F1-weighted scores.Results The constructed multi-classification autism diagnosis model MLAN achieved multi-classification accuracy,F1-macro and F1-weighted scores of 0.850±0.066,0.817±0.103 and 0.834±0.087.The values of all three indicators were better than those of baseline models and the ablated models.Conclusion The proposed MLAN can effectively deal with heterogeneous multi-omics data with small samples and achieve good results,which is expected to provide assistance for the clinical diagnosis of autism.
