Markov Modeling of Side Effect Related Dropout Rates by Introduction of Previous State Memory
Sven Mensing, Peter Nörtersheuser, Paul-Matthias Diderichsen
Abbott GmbH & Co. KG, Ludwigshafen am Rhein (Germany)
Objectives: Describe the relationship between drug-induced level of a common side effect and dropout using a mixed effects markovian exposure-response model. The model assessment was done by predictive checks and external validation.
Methods: A cumulative logistic exposure-response model with markovian elements was developed to describe the observed side effect severity recorded daily for patients in two clinical studies (4 weeks, N= 241 and 6 weeks, N=252). Side effect severity was classified as none (0), mild (1), moderate (2) and severe (3). Dropout was included as a separate state. Drug concentration was assumed to increase transition probabilities to states of more severe side effect. This drug effect was assumed to be reduced by co-medication and tolerance development. A side effect memory boosting the drug effect with a non-linear integration of previous side effect intensities was found to improve the model.
Parameter estimation was performed using NONMEN and predictive checks were performed using simulations with Pharsight Trial Simulator.
Results: The distributions of the state transfer rates from 1000 simulated trial replicates are in good agreement with the observed transfer rates. The 95% confidence interval of the simulations included all observed transition rates, demonstrating the appropriateness of the model.
The dropout observed in a 16 weeks, N=384 study were simulated for external validation using the developed model. Daily recordings of the side effect intensity were not included in this study. The observed dropout rates were enclosed by the 95% confidence interval of the simulation
Conclusions: The Markov model including a side effect memory predicts the observed transition rates and time courses adequately. A Markov model without information of previous side effect intensities was at considerable disparity with the data.
The model was able to predict the drop out rates of a 16 week study based on the data from a 4 week and a 6 week study. The clinical perspective of the modeling is to minimize the dropout as response to side effect intensity by designing/selecting an appropriate run-in period of administered drug and co-medication.
