摘要： 胃癌的早期发现对于降低死亡率、提高患者生存质量具有重要意义。现有预测模型通过单一的结构化电子健康记录数据预测患者的癌症患病风险，但是无法有效地整合不同类型的临床数据且不能满足实际的临床需求。提出一种基于多类型异构数据融合的胃癌风险预测模型。利用预训练语言模型提取电子健康记录数据中的入院记录文本信息，采用降噪自动编码器提取实验室检验数据的特征，同时对低维度的结构化数据向量表示的维度进行扩增，以避免低维度的实验检验特征表示被高维度特征淹没。在此基础上，将扩增后的结构化数据向量与高维度的文本表示向量在相同的尺度上进行融合，从而预测患者的患病风险。实验结果表明，该模型的准确率可达到0.949 337，相比支持向量机、逻辑回归、朴素贝叶斯等模型，具有较优的预测性能。

关键词: 风险预测, 融合模型, 深度学习, 胃癌, 电子病历

Abstract: The early detection of gastric cancer is of great significance for reducing mortality and improving the life quality of patients.Existing prediction models predict the cancer risk of patients using single-structured Electronic Health Record (EHR) data but cannot effectively integrate different types of clinical data or satisfy actual clinical needs. This study proposes a gastric cancer risk prediction model based on multi-type heterogeneous data fusion.The pretraining language model is used to extract text information from the admission record in the EHR data, and the denoising auto-encoder is used to extract features from the laboratory test data.Simultaneously, the dimension represented by the low-dimensional structured-data vector is expanded to avoid drowning the low-dimensional experimental test feature representation by high-dimensional features.On this basis, the expanded structured-data vector and high-dimensional text representation vector are fused on the same scale to predict the patient's disease risk.Experiments show that the accuracy of the proposed model can reach 0.949 337, which is a better prediction performance than that achieved using support vector machine, logistic regression, naive Bayes, and other models.

Key words: risk prediction, fusion model, deep learning, gastric cancer, electronic medical record

中图分类号: 

• TP391.4