A combination of power analysis and machine learning approaches to improve individual prediction of population pharmacokinetic modeling

Objective The purpose of this study is to explore whether combining the power analysis of covariates and the machine learning (ML) analysis of correlated covariates can improve the model’s predictive performance. Methods Two drugs were selected to explore the methodology of population pharmacokinetics. The parameter-covariate relationships were estimated by the SCM and COSSAC methods. The different covariates for each parameter selected by the SCM and COSSAC methods implement the power analysis. And then the highly correlated covariates were analyzed to select a significant covariate by the ML method. The performance of the calibrated model developed by the above process was compared with the performance of two other predictive models: the SCM model and the COSSAC model. Relative error was used to evaluate the different models. The non-compartment analysis (NCA) was used to calculate individual reference AUC. The AUC obtained by population pharmacokinetic models was compared with the reference AUC to evaluate these models’ predictive performance. Result The performance of the calibrated model is similar to the SCM and COSSAC models known from the relative error, the R2of all models is more than 97%. However, the performance predicting the AUC of the calibrated model is better than the SCM and COSSAC models. Conclusion The power analysis can reduce the redundancy of covariate information, and the ML method can provide a criterion to select a covariate from these highly correlated covariates. There combining the two methods may be beneficial for improving individual prediction.