人工神经网络模型在2型糖尿病患病风险预测中的应用

Application of artificial neural network to predict individual risk of type 2 diabetes mellitus

目的：探讨人工神经网络（ ANN）模型在个体2型糖尿病患病风险预测中的应用。方法：通过横断面调查对河南某农村社区8640名居民进行流行病学调查，按3砄1的比例随机分为训练集（6480人）与检验集（2160人），分别用于筛选变量、建立预测模型及对模型的检测和评价。分别应用ANN和logistic回归建立2型糖尿病预测模型，应用受试者工作特征曲线（ ROC）评价预测模型的检验效能。结果：ANN预测模型的灵敏度（95％CI）＝86．93（81．41～91．29）％、特异度（95％CI ）＝79．14（77．18～81．02）％、阳性预测值（95％CI ）＝31．86（28．60～35．03）％、阴性预测值（95％CI）＝98．18（97．37～98．81）％优于logistic回归预测模型[灵敏度（95％CI）＝62．81（55．73～69．47）％、特异度（95％CI）＝71．70（69．52～73．79）％、阳性预测值（95％CI）＝19．94（17．00～22．99）％、阴性预测值（95％CI）＝94．50（93．32～95．57）％]；ANN预测模型AUC（95％CI）＝0．891（0．877～0．905）明显大于logistic回归预测模型[AUC（95％CI）＝0．742（0．722～0．763）]。结论：在预测个体患2型糖尿病方面，ANN模型较logistic回归模型具有更好的预测效能。

Aim：To explore the potential application of artificial neural network （ ANN） on type 2 diabetes mellitus （T2DM）, and then to develop an effective and inexpensive prediction approach .Methods：A cross-sectional survey was conducted.Out of 8 640 subjects who met inclusion criteria, 75%（n1 =6 480） were randomly selected to provide training set for constructing ANN and multivariate logistic regression （ MLR） models.The remaining 25%（ n2 =2 160 ） were as-signed to validation set for performance comparisons of the ANN and MLR models .Predictive performance of different mod-els was analyzed by the receiver operating characteristic （ROC） curve using the validation set.Results：For ANN model, the sensitivity, specificity, positive and negative predictive values for identifying T 2DM were 86.93（81.41-91.29）%, 79.14（77.18-81.02）%, 31.86（28.60-35.03）%, and 98.18（97.37-98.81）%, respectively, while MLR model were only 62.81（55.73 -69.47）%, 71.70（69.52 -73.79）%, 19.94（17.00 -22.99）%, and 94.50（93.32 -95.57）%, respectively.AUC（95%CI） value for identifying T2DM when using the ANN model was 0.891（0.877 -0.905）, showing more accurate predictive performance than the MLR model [AUC（95%CI）=0.742（0.722-0.763）]. Conclusion：The ANN model has a better discriminated performance than MLR model in the prediction of risk of T 2DM.