| 摘要: |
| 目的 构建一种集成神经网络模型实现类风湿关节炎（RA）中医证候分类，并探究其中的特征重要性和风险因素。方法 针对基于人工智能技术的RA中医证候多标签分类中存在的标签关联性差、泛化性能低等问题，提出一种集成神经网络模型——集成神经网络链（FEN）。FEN模型采用一种基于深度神经网络的特征提取基分类器提取临床RA多标签样本的深层特征，增强RA特征区分度；根据协方差理论衡量标签相关性，调节分类器链的输入空间，减少RA错误信息传播和冗余度；并采用集成学习方法减小分类器链中不合理标签序列对RA特征分类的影响。此外，分析了RA中医证候主证和兼证的特征贡献度，挖掘其潜在的风险因素。结果 FEN模型的10折交叉验证性能参数汉明损失、1-错误率、准确度和F1值分别为0.003 6、0.024 8、97.52%、99.18%。与7种典型多标签分类器（分类器链、标签幂集、二进制关联、随机k-标签集、多标签K最近邻、集成分类器链和集成二进制关联）相比，FEN模型具有较好的分类性能。特征贡献度分析提示，主症和次症特征均可作为RA中医证候分类的重要指标，是影响主证和兼证分类的主要因素。结论 基于集成神经网络模型的RA中医证候分类器具有较高的分类精度和效率，对于RA的临床诊断和治疗具有重要参考价值。 |
| 关键词: 类风湿关节炎 多标签学习 神经网络 分类器链 集成学习 |
| DOI：10.16781/j.CN31-2187/R.20230373 |
| 投稿时间：2023-07-02修订日期：2023-12-13 |
| 基金项目:国家自然科学基金（81973749），中国中医科学院科技创新工程重大攻关项目（CI2021A01503）． |
|
| Classifiers for traditional Chinese medicine syndromes of rheumatoid arthritis based on ensemble neural network |
| YANG Jingdong1*,JIANG Biao1,LI Yiwei1,JIANG Quan2,HAN Man2,SONG Mengge2 |
(1. Autonomous Robot Lab, School of Optical-Electrical and Computer Engineering, University of Shanghai for Science and Technology, Shanghai 200093, China;
2. Department of Rheumatology, Guang'anmen Hospital, China Academy of Chinese Medical Sciences, Beijing 100053, China
*Corresponding author) |
| Abstract: |
| Objective To construct an ensemble neural network model to achieve traditional Chinese medicine (TCM) syndrome classification for rheumatoid arthritis (RA) and explore the importance of the features and risk factors. Methods An ensemble neural network model FEN (feature extration network) was proposed to solve the issues such as poor label correlation and low generalization performance in multi-label classification of TCM syndromes of RA based on artificial intelligence technology. The FEN model utilized a feature extraction classifier based on deep neural network to extract deep features from clinical multi-label RA samples, enhancing the discriminative power of RA features. By measuring label correlation based on covariance theory, the input space of the classifier chain was adjusted to reduce the spread of RA error information and redundancy. An ensemble learning method was used to mitigate the impact of unreasonable label sequences in the classifier chain on RA feature classification. Additionally, the importance of main and accompanying TCM syndrome features of RA was analyzed and potential risk factors were explored. Results The FEN model had excellent performance in a 10-fold cross-validation, with Hamming loss, one-error, accuracy, and F1-score being 0.003 6, 0.024 8, 97.52%, and 99.18%, respectively. Compared with 7 typical multi-label classifiers (classifier chain, label powerset, binary relevance, random k-labelsets, multi-label K-nearest neighbor, ensemble classifier chain, and ensemble binary relevance), the FEN model had better classification capabilities. The analysis of feature contribution indicated that the features of main and secondary symptoms might be used as important indicators of classification of TCM syndromes of RA, and were the main factors affecting the classification of main and accompanying syndromes. Conclusion The RA TCM syndrome classifier based on ensemble neural network has high classification accuracy and efficiency, providing important reference for the clinical diagnosis and treatment of RA. |
| Key words: rheumatoid arthritis multi-label learning neural network classifier chain ensemble learning |
