遗传算法优化BP神经网络的异丙酚血药浓度预测

Plasma Concentration Predication for Propofol of Optimized BP Neural Network Based on Genetic Algorithm

针对静脉麻醉异丙酚时变性强,房室结构复杂的特性,经典的非线性混合效应模型参数估计法存在变量繁多,人为因素多等弊端,而BP神经网络又存在极易陷入局部极值,网络训练不稳定致使预测误差大。利用遗传算法优化BP神经网络的权值和阈值,调整神经网络中异丙酚血药浓度和时间、病人年龄、体重、身高、体表面积、采样时间、总剂量、注射率的关系,然后建立异丙酚血药浓度预测模型,并与NONMEM方法、BP神经网络进行比较。比较结果,GA-BP网络的平均误差为1.2%,BP网络的平均误差为29.59%,NONMEM为14.61%,GA-BP网络的绝对平均误差15.76%,BP网络的绝对平均误差31.9%,NONMEM为22.99%。实验结果表明:GA-BP网络对于半衰期较短的麻醉药物异丙酚药物具有较好的非线性拟合能力和更高的预测准确性。

Due to the nature of propofol of high time-varying, and complex compartment model, the traditional estimation method of Nonlinear Mixed Effects Model（NONMEM） has miscellaneous of variables and plenty of man- made jamming factors while BP neural network is easy to fall into the local extremum, and unstable network train- ing. This study used GA to optimize the weights and thresholds of BP neural network, and adjusts the relationships between the patients＂ plasma concentration of propofol and time, patient＇ age, gender, lean body mass, height, body surface area, sampling time, total dose, and injection rate through network training, to build a model of pre- diction the plasma concentration of propofol, and after that, it compares them with the results of NONMEM algo- rithm and BP neural network. In conclusion, the average error of GA-BP neural network is 1.20% , while that of BP neural network and NONMEM and is 29.59% , 15.76% , respectively. The absolute average error of GA-BP neural network is 15.76%, while that of BP neural network and NONMEM and is 31.9%, 22.99%, respectively. The experimental results indicated that the proposed method is suitable to fitting of short half-life period anesthesia drug propofol with higher predication accuracy.