基于径向基函数神经网络的HIV感染者的识别预测

A HIV carrier recognition model based on radial basis function neural network

目的应用径向基函数神经网络,在描述性分析的基础上,建立艾滋病病毒(HIV)感染者的预测模型,分析预测效果。方法在描述性分析的基础上,以患病种类为因变量,以年龄、国籍、文化程度、职业、劳务史、会阴部症状史、不洁性生活史、明确性伴侣有性病、是否有同性性伴侣、合法性伴侣是否是高危人群、性伴侣数为自变量,建立径向基函数神经网络,分析预测效果和预测变量的重要性。结果选择2004-2008年某省口岸出入境人员体检监测中检出的HIV感染者、梅毒患者和非性病人员各98例,HIV感染者以35～49岁男性为主(67.35%),中学以上文化程度占81.63%,职业以劳务(46.94%)和公务(30.61%)为主,有2例外籍女性性服务工作者和2例外教。径向基函数神经网络模型,对训练样本和检验样本的预测总准确率为85.37%和85.00%。用该模型对独立样本进行预测,总准确率为89.66%,对HIV感染者、梅毒患者和非性病患者的准确率分别为83.33%、85.71%和100.00%。三类预测结果ROC(Receiver operating characteristic)曲线下面积都>0.90,年龄、性伴侣数和劳务史是最重要的三个影响因素。结论口岸HIV感染者有其特定的流行病学特征,HIV感染者径向基函数神经网络预测模型的拟合能力强,能较好地用于评价未知样本。

Objective To establish a HIV carrier recognition model based on radial basis function neural network and analyze the effectiveness of this model. Methods The epidemiologic feature of 98 HIV carriers found in ports of one Chinese province from 2004 to 2008 was analyzed. Another 98 syphilis patients and 98 people without sexual transmitted disease were also analyzed. First, the epidemiological features of the HIV carriers were analyzed. Second,the recognition model based on radial basis function neural network for HIV carriers was established. The independent variables included： age, nationality, educational level, occupation, history of labor service, manifestation of perineal symptom, unprotected sexual activity, sex partners with STD, homosexual partners, regular sex partners at high risk for STD, number of sex partners. Third, the effect of forecasting model and the importance of all independent variables were analyzed. Results The majority of 98 HIV carriers were male, who were at the age from 35 to 49 and had middle school or higher education. The main occupations were labor （46.94%） and public service （30.61%）. Two of them were foreign female commercial sex workers, and other two were foreign teachers. The overall forecasting accurate rate of this model for total training sample and total test sample were 85.37% and 85.00%, respectively. For another independent sample, overall forecasting accurate rate was 89.66 %, and accurate rate for HIV carriers, syphilis patients and people without STD were 83.33% , 85.71% and 100.00% ,respectively. Area under the ROC curve for there different classifications all exceeded 0.90. Age, number of sex partner and history of labor service were the most important impact factors. Conclusion HIV carriers found in ports had their specific epidemiological features and the model of HIV carriers recognition based on radial basis function neural network had good fitting ability,which could be used for predicting unknown sample effectively.