Predictive assessments of pharmacokinetic alterations in subjects with renal disease
Elisa Borella (1), Italo Poggesi (2), Paolo Magni (1).
(1) Dipartimento di Ingegneria Industriale e dell'Informazione, Universitŗ degli Studi di Pavia, Italy; (2) Quantitative Sciences/Model-Based Drug Development, Janssen-Cilag SpA, Italy.
Objectives: To find a model capable of predicting, relying on a minimum amount of PK information in normal subjects, the effect of renal impairment on the exposure of a drug. Three categories of renal impairment (mild, moderate and severe) were considered according to the KDIQO guideline.
Methods: For a list of 64 marketed drugs, PK descriptors and recommended dosage adjustments for subjects with renal impairment were obtained from the literature and other public resources . The considered PK descriptors were: clearance, bioavailability, oral clearance, amount excreted unchanged in urine (Ae), binding to plasma protein (ppb) and hepatic extraction ratio. The PK changes due to the renal dysfunction were summarized using the ratio between the AUC computed in subjects with renal impairment and the AUC in healthy subjects. Two different types of analyses were performed. The first one consisted in exploring, for each level of renal impairment, the potential correlations between the AUC ratios and the PK variables, implementing in R different methods, from multiple linear regression (MLR) to partial least squares (PLS). The second approach consisted in classifying the drugs, basing always on their PK parameters, in different levels of risk of administering wrong dosage to renal impaired patients. Different data-mining and machine-learning methods were tested using Orange 2.7 .
Results: A multiple regression (regressors selected by a stepwise procedure) gave results with a reasonable degree of accuracy. Besides, the regressors selected (Ae for mild and moderate, Ae and ppb for severe renal impairment) find a physiological explanation. The results of the regression analysis were summarized in a sort of “traffic-light” histograms which could become an immediate clinical decision-making tool. As regards the classification problem, for a binary class (risk and no-risk) the best results in term of accuracy were given by Naïve Bayes, Classification Trees and SVM. Setting a threshold to discern risky and not-risky administering situation, the AUC predicted by the MLR were discretized and the percentage of misclassifications was compared to those produced by the best classificators.
Conclusions: This model can be considered a smarter statistical analysis, which may provide a useful guidance for designing the studies of new compounds, highlighting those cases that may need additional investigations.
Acknowledgements: This work was supported by the DDMoRe project (www.ddmore.eu).