Predictability and identifiability assessment of models for prostate cancer under androgen suppression therapy

The past two decades have seen the development of numerous mathematical models tostudy various aspects of prostate cancer in clinical settings. These models often contain large sets ofparameters and rely on limited data sets for validation. The quantitative analysis of the dynamics ofprostate cancer under treatment may be hindered by the lack of identifiability of the parameters fromthe available data, which limits the predictive ability of the model. Using three ordinary differentialequation models as case studies, we carry out a numerical investigation of the identifiability and uncer-tainty quantification of the model parameters. In most cases, the parameters are not identifiable fromtime series of prostate-specific antigen, which is used as a clinical proxy for tumor progression. Itmay not be possible to define a finite confidence bound on an unidentifiable parameter, and the relativeuncertainties in even identifiable parameters may be large in some cases. The Fisher information ma-trix may be used to determine identifiable parameter subsets for a given model. The use of biologicalconstraints and additional types of measurements, should they become available, may reduce theseuncertainties. Ensemble Kalman filtering may provide clinically useful, short-term predictions of pa-tient outcomes from imperfect models, though care must be taken when estimating “patient-specific”parameters. Our results demonstrate the importance of parameter identifiability in the validation andpredictive ability of mathematical models of prostate tumor treatment. Observing-system simulationexperiments, widely used in meteorology, may prove useful in the development of biomathematicalmodels intended for future clinical application.