2.2.2 TGI-OS Modeling
The TGI-OS model is a disease progression model. It is a useful tool in oncology to delineate E-R relationships in the presence of confounding factors. The model is composed of two parts: a TGI model that describes tumor dynamics, and a multivariate survival model that incorporates a TGI metric as a covariate on OS. The TGI metric serves as a marker of disease status. TGI-OS modeling mitigates confounding in the E-R analysis by directly evaluating the treatment effect on TGI then separately accounting for the effect of prognostic factors on OS. By mitigating the confounding effects of prognostic factors on the relationship between treatment effect and OS, the TGI-OS model can avoid a false positive E-R relationship.24-26 The TGI model structure is typically a simple biexponential model (Equation 3).27
\(f\left(t\right)=exp\ \left(-d\times t\right)\ +exp\ \left(g\times t\right)\ -1\)(Equation 3 ) 27
where f(t) is tumor size at time t , d is the decay rate constant, and g is the growth rate constant.
In multiple cancer types, the OS is correlated with the tumor dynamics such that the probability of survival decreases with the increase in tumor growth rate (g in Equation 3).27-37 Other key determinants for survival are baseline prognostic factors specific for the cancer type. Drug exposure is evaluated as a covariate in the multivariate survival model.24, 25, 29, 32 If it is not significant this suggests a flat E-R relationship. For exposure-driven TGI models drug exposure is not evaluated as a covariate in the multivariate survival model. OS can be simulated for exposure quartiles with normalized prognostic factors to evaluate the presence of an E-R relationship. This approach can remove the confounding effects of imbalanced prognostic factors in different exposure quartiles. TGI-OS modeling has successfully evaluated E-R relationships for atezolizumab in multiple indications, and its role in E-R analysis has been increasingly accepted by regulatory agencies.25, 28, 29, 32
While TGI-OS modeling allows for the direct separation of treatment effect and disease effects, several limitations must be considered. Non-exposure driven TGI models while simpler and more flexible than exposure-driven models require assumptions and empirical descriptions of tumor shrinkage and growth. Model building for both exposure and non-exposure driven models requires one or more post-treatment assessments for tumor size, and this may not be feasible in some patients. The incorporation of multiple tumor size assessments in the model, however, makes tumor dynamics a patient-specific explanatory variable and informative predictor of survival. With the TGI-OS model, it is also difficult to account for the potential appearance of new lesions. Zecchin et al . developed a pharmacometric model to incorporate the effect of new lesions on OS in metastatic ovarian cancer, but additional examples and uses of this approach are currently limited.38, 39 Because the TGI-OS model predicts OS based on tumor dynamics it is more suitable for use in advanced malignancies, where tumor size is typically measured over time.