Evaluation of individual probability in mixture model in NONMEM®
Other conference paper



摘要Objectives: One of the utilities of mixture model in NONMEM® is to estimate the probability of a subject belonging to each subgroup (PMIX) and classify the subject according to the highest value (PMIXmax), based on its pharmacokinetic data. Theoretically, the probability of correct classification (i.e. the reliability of the classification), Pcorrect, is estimated by PMIXmax, but it has not been investigated before. This study aims at evaluating PMIXmax against the observed Pcorrect and providing a better estimator for Pcorrect if necessary.

Methods: 59,049 pharmacokinetic datasets were simulated using the subroutine ADVAN2 TRANS2 with a mixture of two CL distributions, with varying parameters (including sample size, typical values and variability of CL, Vd and ka, mixing proportion and residual unexplained variability). A model with the $MIX block and another without were run on each dataset to obtain the estimates for PMIXmax for each subject, and the change in OFV per observation (DOFVobs) after removing the mixture model, respectively. Data were binned by DOFVobs and PMIXmax, and Pcorrect in each bin was computed and compared against PMIXmax. The relationship between Pcorrect and PMIXmax and DOFVobs would be investigated and an empirical equation to estimate Pcorrect would then be developed and evaluated.

Results: The estimation of Pcorrect by PMIXmax was poor. In terms of binned data, PMIXmax almost always overestimates Pcorrect (R2=0.693, MPE=+7.96%), especially when DOFVobs is small (Figure). An estimation equation for Pcorrect based on DOFVobs and PMIXmax was then developed, showing improved performance (R2=0.952, MPE=+0.02%).

Conclusions: PMIXmax is not a reliable estimator for Pcorrect. The empirical estimation equation for Pcorrect based on DOFVobs and PMIXmax greatly improves the estimation. This study has confirmed the biases of PMIXmax and demonstrated the possibility to improve the estimation of Pcorrect.
著者HUI Ka Ho, LAM Tai Ning
會議名稱The 9th American Conference on Pharmacometrics
會議地點San Diego

上次更新時間 2019-20-06 於 15:25