医学知识增强的肿瘤分期多任务学习模型

肿瘤分期是指从病人的电子病历文本中推测肿瘤对应阶段的过程。在电子病历数据中存在类别严重不均衡现象，因此使用深度学习方法进行肿瘤分期具有一定的挑战性。该文提出医学知识增强的多任务学习KEMT(knowledge enhanced multi-task) 模型，将肿瘤分期问题视作面向医疗电子病历的文本分类任务，同时引入医生在人工预测肿瘤分期时参考的医学属性，提出基于医学问题的机器阅读理解任务，对上述两种任务进行联合学习。我们与医疗机构合作构建了真实场景下的肿瘤分期的数据集，实验结果显示，KEMT模型可以将医学知识与神经网络结合起来，预测准确率高于传统的文本分类模型。在数据分布不均衡的条件下，在小样本类别上的准确率提升了4.2个百分点，同时模型也具有一定的解释性。

Multi-task tumor stage learning model with medical knowledge enhancement

Tumor staging is the process of inferring the corresponding stage of tumors based on patients’ electronic health records (EHR). The serious uneven data distribution in the types of EHRs has certain challenges on tumor stage prediction through in-depth learning. Accordingly, this paper proposes a knowledge enhanced multi-task (KEMT) model and considers tumor stage reasoning as a text classification task of EHR. It also introduces medical attributes that doctors referred to in tumor stage prediction and introduces a medical problem-based machine reading comprehension task. The tasks are jointly studied by building a real-world dataset of tumor staging with medical institutions. Experimental results show that the KEMT model combines medical knowledge with a neural network and gets a higher precision rate of prediction than the traditional text classification models. Under the condition of uneven data distribution, the accuracy of small samples is improved by 4.2%, for which the model also accounts.