摘要
针对医疗数据的智能化识别与分析需求,文中对医疗财务大数据挖掘的相关方法进行了研究。通过引入深度学习中的深度置信网络(DBN),结合Autoencoder自编码网络构建了数据处理系统,实现对医院经营状态的自动化评估。DBN网络使用受限玻尔兹曼机(RBM)替代了传统神经网络中神经元结构作为网络的隐藏层,该结构可以多个堆叠,提升网络的泛化能力。使用Gibbs抽样,得到RBM的近似分布,提升算法的训练效率。同时Autoencoder网络可以从大维度的财务经营数据中,筛选出更能描述数据特性的特征维度。为了验证系统算法的性能,在某医院的财务数据集上进行测试,使用Autoencoder自动提取17个财务数据指标作为模型的输入特征,以评估结果作为模型的输出向量。对比实验结果表明,相较于逻辑回归、BP神经网络等浅层的机器学习算法,文中算法的AUC与Accuracy分别可以达到0.81、80.0%,具有较为明显的提升。
According to the needs of intelligent identification and analysis of medical data,The related methods of medical financial big data mining are studied.By introducing the deep confidence network(DBN)in the deep learning and combining with the Autoencoder self coding network,the data processing system is constructed to realize the automatic evaluation of the hospital operation status.The Restricted Boltzmann Machine(RBM)is used in DBN instead of the traditional neural network as the hidden layer of the network.The structure can be stacked in multiple layers to improve the generalization ability of the network.Using Gibbs sampling,the approximate distribution of RBM is obtained,which improves the training efficiency of the algorithm.At the same time,the Autoencoder network can filter out the feature dimensions which can describe the data characteristics better from the large-scale financial operation data.In order to verify the performance of the system algorithm,the test is carried out on the financial data set of a hospital.The test uses Autoencoder to automatically extract 17 financial data indicators as the input characteristics of the model,and the evaluation results as the output vector of the model.The experimental results show that the AUC and accuracy of the algorithm can reach 0.81 and 80.0%respectively,compared with the shallow machine learning algorithms such as logic regression and BP neural network,which has a significant improvement.
作者
谷丽霞
刘欣芃
GU Lixia;LIU Xinpeng(Shanghai Sixth People’s Hospital,Shanghai 201303,China;Sias University,Zhengzhou University,Xinzheng 451150,China)
出处
《电子设计工程》
2021年第10期46-50,共5页
Electronic Design Engineering
基金
上海市科技厅项目(3160235B)。
