Study On Radial Basis Function Neural Network Model And Intelligent Expert System In Population Pharmacokinetics

[Objective] Population pharmacokinetics can quantitative describe the variance of the patient population's pharmacokinetic parameters by applying the classical pharmaco- kinetics fundamental principles and statistical methods. It is useful for guiding the clinical medication. The traditionary methods to estimate population pharmacokinetics parameters must depend on the existing models, which has the objections of heavy workload and many jamming factors. It is likely to hide some risk to guide the clinical medication by applying these models. Moreover, the mainstream population pharmacokinetics analysis software, NONMEM program, is very expensive, resulting in domestic population pharmacokinetics study developed slowly. Artificial neural network does not depend on the existing model and can find out the essential rule implied in the data by itself, according to its input and output parameters. Some foreign researchers have studied population pharmacokinetics using artificial neural network. The used artificial neural network is BP neural network mostly. BP neural network has some defects, such as complicated structure, many required definited parameters, converge slowly, fall into a local minimum easily and so on. A relatively new artificial neural network, radial basis function neural network, has some fortes, such as simple structure, few parameters required determined, converge fast, fall into a local minimum hardly and so on. In this study, radial basis function neural network was used to analyze population pharmacokinetics, establish steady plasma-drug concentration predictive models and carried out the dose titration to obtain the patients'best dosage regimen. The full process was implemented by a series of programs compiled by myself. The objective of the study was to investigate the feasibility of application of radial basis function neural network to population pharmacokinetics analysis and production of intelligent expert system.[Method] In this study, applied radial basis function neural network, which compiled by MATLAB 7.0 software, to analyze the analog data, which has different sample size, and the information of some patient populations who were treated with perphenazine, risperidone, clozapine, sulpiride or chloropromazine. And adjusted the dose of the patient whose steady plasma-drug concentration was out of effective plasma-drug concentration range to cause the steady plasma-drug concentration be in the range according to the five steady plasma- drug concentration predictive models. Furthermore, merged the five drug data and established the model to predict the five drugs'steady plasma-drug concentration predictive models. Evaluated the model learning effect with error of mean square (MSE) and coefficient correlation (R) between the computed output value and objective output value of training set and validation set, respectively. Evaluated the model predict performance with error of mean square (MSE) and coefficient correlation (R) between the computed output value and objective output value of test set.[Result] Radial basis function neural network converged fast very much, the eplased time was under 100s to modeling data which sample number less than 1000, while the eplased time was under 20s to modeling data which sample number less than 200 in this study. The training set MSE-values of Net1000, Net500 and Net100 model with different sample size were beyond 10^-6, while R-values were 0.9842, 0.97353 and 0.97122, respectively. And the test set MSE-values of Net1000, Net500 and Net100 were 0.13541, 0.17876 and 0.72051, while R-values were 0.98006, 0.97630 and 0.87984, respectively. So, it is reasonable to comprehensive evaluate the model learning effect combinated training set and validation set. Bigger the training set sample size, more fully the network learns and better the model performance. The test set MSE-values between computed steady plasma-drug concentration and observed steady plasma-drug concentration of perphenazine, clozapine and chloropromazine were 0.016923,0.005439 and 0.0016704, while R-values were 0.87635,0.93676 and 0.98265, respectively. The predict performance of perphenazine, clozapine and chloropromazine models is satisfactory. The models are used to adjust the drug dose for guide the clinical medication. The test set MSE-values between computed steady plasma-drug concentration and observed steady plasma-drug concentration of risperidone and sulpiride were 0.00858 and 0.011001, while R-values were 0.80899 and 0.81425, respectively. The model is immature yet. The corresponding MSE-value and R-value of the model Net containing five drugs were 0.012292 and 0.8896, which is satisfactory.[Conclusion] Radial basis function neural network has some characteristic that simple structure, convergence fast , and self-learning and self-adaptation. It can further improve the network models performance and helpful for guiding the clinical medication by means of keeping collecting population pharmacokinetics sample information of perphenazine, risperidone, clozapine, sulpiride and chloropromazine and making the NetI-NetV models on-line learning endlessly. And it is possible to making a intelligent expert system for predicting many kinds of drugs'steady plasma-drug concentration by means of keeping collecting all kinds of drugs'population pharmacokinetics sample information and making Net model on-line learning endlessly. It has enormous potential prospect that radial basis function neural network apply to population pharmacokinetics analysis. Radial basis function neural network would bring into play important role in population pharmacokinetics field. It can improve the drug efficacy and reduce the occurrence of toxicity reaction and adverse reaction, if radial basis function neural network was used to guide clinical medication. Radial basis function neural network would become a powerful tool for medicine workers.