BP神经网络在预测血液病患者医院感染中的应用

BP neural network in prediction of nosocomial infections in patients with hematological diseases

目的通过分析成年血液病患者的数据,建立BP神经网络模型对患者发生医院感染进行预测。方法分析5 555例成年血液病患者的临床资料,按有无感染分为两组,结合主要指标建立BP网络预测模型,并进行模型验证,检测网络的稳定性,于[0,1]内间隔0.01选取阈值,对模型进行检验,并与传统logistic回归模型预测结果进行对比。结果神经网络模型检测阈值<0.39时,随阈值的增大总准确率逐渐升高;阈值>0.39时,总准确率随阈值的增大呈下降趋势,以0.39为阈值,测试样本预测结果,敏感度为0.520,特异度为0.870,阳性预测值为0.371,阴性预测值为0.924,与实际感染情况比较,BP神经网络模型医院感染预测准确率为82.49%,高于多元logistic回归模型的预测结果74.13%。结论利用BP神经网络模型对血液病患者医院感染预测有良好的效果。

OBJECTIVE To establish BP neural network to predict nosocomial infections in patients with hematological diseases through analysis of the data of the adult patients with hematological diseases. METHODS The clinical data of 5555 adult patients with hematological diseases were analyzed, the subjects were divided into two groups, then the BP neural network predication model was established by combing with the key indicators, the model was verified, and the stability of the network was tested; the threshold value was determined with an interval of 0.01 in [0,1], the model was tested, and the test result was compared with the prediction result of traditional logistic regression model. RESULTS When the threshold value of the neural network model test was less than 0.39, the accuracy rate was gradually increased with the increase of the threshold valuer when the threshold value was more than 0.39, the accuracy rate showed an downward trend with the increase of threshold value; as 0.39 was defined as the threshold value for the predication of tested samples, the sensitivity was 0. 520, the specificity was 0. 870, the positive predication value was 0. 371, and the negative predication value was 0. 924; as compared with the actual incidence of infections, the accuracy rate of BP neural network model in predication of nosocomial infections was 82.49M, higher than 74.13% of the multivariate logistic regression model. CONCLUSION The BP neural network model can achieve good effect on prediction of nosocomial infections in the patients with hematological diseases.