在糖尿病危险因素分析中人工神经网络的应用

目的 研究用神经网络(NN)分析疾病危险因素。方法 以糖尿病流行病学调查资料为基础,用BP网作为拟合模型,网络的结构为22-6-1,分析糖尿病各种可能危险因素的平均影响值(MIV),按MIV值的绝对值大小排出因子顺位,并与logistic回归模型的分析结果相比较,用对数线性模型分析因子间的联合作用。结果 NN多因素分析,糖尿病危险因子顺位为脉搏、糖尿病家族史、居住年限、肾病史、腰臀比、性别、高脂血症史、冠心病史、高血压病史、收缩压、收入、饮酒、年龄、舒张压、文化程度、体重指数、其他病史、职业性体力活动、吸烟、职业、脑血管病史、肝病史;而多因素logistic回归分析中只有7个变量入选最终模型,因子顺位为脉搏、糖尿病家族史、肾病史、腰臀比、高血压病史、职业性体力活动及年龄;对数线性模型分析发现二者的差别可由因子间的相互作用解释一部分。结论 NN完全能够胜任疾病危险因素的分析任务,可拟合出比传统模型更复杂的变量间关系。

Study on the application of artificial neural network in analysing the risk factors of diabetes mellitus

Objective To study the use of neural network in determining the risk factors of diseases, Methods With back-propagation neural network (BP network) as fitting model based upon data gathered from an epidemiological survey on diabetes mellitus and under the network structure of 22-6-1, the mean impact value (MIV) for each input variables and sequencing the factors according to their absolute MIVs were calculated. The results from BP network with multiple logistic regression analysis and log-linear model for united actions between factors were compared with optimizing Levenberg-Marquardt algorithm. Results By BP network analysis, the sequence of importance for the risk factors of diabetes mellitus became: faster pulse, diabetes mellitus family history, living longer in the investigated area, with medical record of nephropathy, having higher ratio for waist-to-hip, being male, with medical records of diseases as hyperlipoproteinmia, coronary heart disease, hypertension, high diastolic pressure, higher income, do no drink alcohol, age, higher systolic pressure, less educated, body mass index, with medical records of other diseases, physieal exercise related to jobs smoking, occuppation, with medical record for cerebrovascular disease, with medical record for liver disease etc. However,only 7 factors were statistically significant in multiple logistic regression analysis. The sequence of their importance appeared as: pulse, diabetes mellitus family history, the medical record of nephropathy, waist-to-hip ratio, the medical record of hypertension, work-place related exercise and age. The sequences of importance were almost the same between the two while the difference could partly be explained by the interaction among risk factors through log-linear model. Conclusion Neural network could be used to analyze the risk factors of diseases and could assimilate more complicated relationships ( main effects and interactions) between inputs and outputs, better than using the traditional methods.