融合医学本体知识的药物推荐算法

Drug recommendation algorithm integrating medicalontology representation

针对传统的药物推荐技术忽略了医疗编码体系中蕴含的本体分类信息,导致推荐效率不高的问题,提出了一种融合图神经网络技术和注意力机制的药物推荐算法。首先通过图神经网络学习医学本体图中的分类关系,对医疗代码进行嵌入表示;其次将医疗表示输入结合注意力机制的循环神经网络中捕捉患者病历信息特征,同时引入药物相互作用知识,通过图神经网络学习药物相互作用关系;最后进行多标签分类来输出推荐药物。算法在电子病历数据集上进行试验,结果如下:F_(1)值为63.09%,杰卡德相似系数为47.43%,精确度调用曲线值为71.64%,优于对比的其他方法。试验结果表明本算法能够丰富医疗代码表示,提升药物推荐的准确率,并且降低推荐药物组合中相互作用概率。本算法能够向医生推荐合适的药物,对医疗智能化发展具有一定的参考价值。

In response to the low efficiency of recommendation arising from the problem that the traditional drug recommendation technology tends to ignore the medical ontology information contained in the medical code,a new drug recommendation algorithm was proposed by integrating the graph neural network with the attention mechanism.Firstly,the algorithm applied the graph neural network technology to learn the classification relationship in the medical ontology and conducted the embedded representation of medical code.Secondly,the patient s medical record information characteristics were captured through recurrent neural networks in combination with the attention mechanism.In addition,the knowledge of drug interaction was introduced to learn the drug interaction relationship through the graph neural network.Finally,the recommended drugs were output through the multi-label classification.The algorithm was tested on an EMR data set,whose experimental results show that the F_(1) value is 63.09%,the Jaccard similarity coefficient is 47.43%,and the accurate call curve is 71.64%,which proves this method is superior to the contrast methods.The algorithm can enrich the representation of medical code,improve the accuracy of drug recommendations,and reduce the interaction rate in the recommended drug combination.The algorithm can recommend the appropriate drugs for doctors,which is of reference value for the development of medical intelligence.