反向传播神经网络在监测颅内压数学模型中的应用

The application of back-propagation neural networks in mathematical model of intracranial pressure

目的探讨反向传播神经网络(BPNN)在建立无创颅内压监测数学模型中应用的可能性及在该领域的应用前景。方法对10例颅脑外伤所致急性颅高压患者连续20min采集其大脑中动脉平均血流速度(VMCA)、平均动脉压(MAP)、呼气末二氧化碳分压(PETCO2)和心率,同时在硬膜外置探头监测颅内压,共获2911组数据。通过Matlab7.0软件中的神经网络工具箱,建立众参数和颅内压的贝叶斯正规化3层BPNN预测模型,进行训练样本和预测样本的仿真模拟,并计算出各个因素的平均影响值(MIV)。结果BPNN模型结构为4-20-1,训练至191步时网络收敛。预测样本的预测值和真实值的相关系数r=0.99,平均绝对误差为1.17mmHg,平均相对误差为7.36%。按照MIV绝对值大小列出4个因素,对于颅内压影响的相对重要性顺位为VMCA、PETCO2、MAP和心率。结论反向传播神经网络模型与颅高压的非线性特性相契合,对颅内压的预测效果良好,能较好地处理颅高压内部多因素间复杂的非线性关系。

Objective To explore the possiblity of the application of back propagation neural networks （BPNN） in establishing mathematical model of non-invasive intracranial pressure monitoring and its application prospect in this field. Methods Mean velocity of middle cerebral artery （ VMCA ）, mean arterial pressure （ MAP）, end-tidal carbon dioxide （ PETCO2 ） and heart rate （HR） were collected continuously for 20 minutes in 10 patients with acute intracranial hypertension （ICH） caused by craniocerebral trauma. Intracranial pressure was monitored with an epidural sensor at the same time, and 2911 data were obtained altogether. Three layers BPNN prediction model of Bayesian regularization in all parameters and intracranial pressure was established by Neural Network Toolbox in Matlab 7.0 software. The emulation simulation of training and prediction samples were performed, and the mean impact value （MIV） of all factors was caculated. Results The structure of BPNN model was 4-20-1. The training did not stop until the networks are convergent at 191 epochs. The correlation coefficient of predictive value and true value of prediction samples was R =0.99, and the mean absolute error and fractional error were 1.17 mm Hg and 7. 36%, respectively. Four factors were listed according to the order of the absolute value of MIV, and the order of their relative importance was VMCA, PETCO2 , MAP, and heart rate. Conclusion The model of BPNN is correspond to the non-linear characteristic of intracranial hypertension. It has a favorable effect on the predition of intracranial pressure, and it may preferably deal with the complicated non-linear relations between factors within intracranial hypertension.