多层感知器神经网络模型对职业性煤工尘肺发病预测研究  

Prediction of occupational coal workers′ pneumoconiosis incidence by a multi-layer perceptron neural network model

作　　者：肖淑玉[1] 高静[1] 孙志谦 汪鹏 张旭 王健龄 沈福海 XIAO Shu-yu;GAO Jing;SUN Zhi-qian;WANG Peng;ZHANG Xu;WANG Jian-ling;SHEN Fu-hai(Tangshan Center for Disease Control and Prevention,Tangshan,Hebei 063000,China)

机构地区：[1]唐山市疾病预防控制中心,河北唐山063000 [2]华北理工大学公共卫生学院,河北省煤矿卫生与安全实验室,河北唐山063210

出　　处：《中国职业医学》2021年第1期19-25,共7页China Occupational Medicine

基　　金：河北省教育厅基金(ZD2018226)。

摘　　要：目的构建并验证基于多层感知器(MLP)神经网络的煤矿接触粉尘(以下简称"接尘")工人职业性煤工尘肺(CWP)发病预测模型,探讨其对CWP发病预测中的应用价值。方法采用典型抽样方法,以河北省某煤矿集团1970—2017年开始从事煤矿开采工作的17 023名接尘工人为研究对象。其中,罹患CWP者839例,未罹患CWP者16 185人。以研究对象CWP发病与否为目标输出变量,工种、年龄、开始接尘年代、观察年(即潜伏期)和累积接尘量为输入变量,建立MLP神经网络模型,采用受试者工作特征(ROC)曲线对所建模型的预测能力进行评价。采用所建立的模型预测未来10年接尘工人CWP发病高危人群和重点监护人群。结果所建立MLP神经网络模型隐含层有44个神经元突触;ROC曲线下面积为0.91,模型准确度为92.7%,灵敏度为74.8%,特异度为93.6%。采用验证样本进行模型验证,准确度为92.1%,灵敏度为70.5%,特异度为93.2%。采用该MLP神经网络模型进行预测,该煤矿集团接尘工人未来10年内发生CWP的高危人群1 534例,并可定位到个体;需要重点监护的危险人数为7 599人。其中,预测不同工种接尘工人未来10年CWP的发生率由高到低依次为掘进工、采煤工、混合工和辅助工(P<0.01),开始接尘年代越早者发生CWP的风险越高(P<0.01)。结论基于工种、年龄、开始接尘年代、潜伏期和累积接尘量构建的MLP神经网络模型应用于煤矿接尘工人CWP发病预测具有较好的效能,可为早期采取预防性管理措施防治CWP提供参考。Objective To construct and verify the incidence prediction model of occupational coal workers′ pneumoconiosis(CWP) in coal mine workers exposed to dust(hereinafter referred to as ″dust exposure″) based on a multi-layer perceptron(MLP) neural network, and explore its application value in predicting CWP incidence. Methods A total of 17 023 dust exposed workers in a coal mining group in Hebei Province from 1970 to 2017 were selected as the research subjects by a typical sampling method. Among them, 839 patients were confirmed as CWP and 16 185 workers did not suffered from CWP. The MLP neural network model was established with the incidence of CWP as the target output variable, and the type of work, age, beginning year of dust exposure, observation year(i.e. incubation period) and cumulative dust exposure as the input variable. The receiver operating characteristic(ROC) curve was used to evaluate the predictive ability of the built model. The established model was used to predict the high-risk group and key monitoring group population of CWP in dust-exposed workers in the following 10 years. Results There were 44 synapses in the hidden layer of the established MLP neural network model. The area under ROC curve was 0.91. The accuracy, sensitivity and specificity of the model were 92.7%, 74.8% and 93.6%, respectively. In the validation samples, the accuracy, sensitivity and specificity were 92.1%, 70.5% and 93.2%, respectively. The MLP neural network model was used to predict 1 534 workers with high risk of CWP in the following 10 years, and the individuals were located. The number of workers in need of actively monitored was 7 599. Among them, it is predicted that the incidence of CWP in different types of dust exposed workers in the following 10 years from high to low is tunneling worker, coal miner, mixing worker and auxiliary worker(P<0.01). The earlier the dust exposure began, the higher the risk of CWP(P<0.01). Conclusion The MLP neural network model based on the type of work, age, beginning year of dust e

关 键 词：煤工尘肺 粉尘 神经网络 发病 预测 模型 高危人群 

分 类 号：R135.2[医药卫生—劳动卫生]