Semi-parametric modelling for analysing the evolution of toxicities over time

In oncology clinical research, reporting adverse events (AEs) is absolutely essential to depict the toxicity profile of new therapeutics such as targeted therapies or immunotherapies. However, conventional methods of analysis are unable to capture the new patterns of AEs induced by innovative drugs with prolonged exposure times. An efficient approach to deal with time-to-onset, duration and the recurrent nature of AEs is to use the prevalence function modelling the evolution of toxicities over time. This non-parametric method proposed by Pepe et al., however, does not consider confounding factors potentially affecting the probability of experiencing AEs such as patients’ clinico-pathological characteristics. To take these potentially covariates of interest into account, one solution is to use a semi-parametric approach. An adaptation of the prevalence function using Cox’s methodology was developed. To illustrate its value, this new approach was applied to data from the MOTIVATE trial, and hypothetical examples were simulated in several scenarios varying time-to-onset, duration and recurrent nature of AEs in presence of a confounding factor (ECOG). Applied to MOTIVATE trial, the semi-parametric prevalence gave similar results to non-parametric one, with the same trend being observed for each clinical parameter. In the first three simulated scenarios, more patients with ECOG > 1 were treated with treatment B, and unadjusted prevalence was significantly higher for treatment B. After adjusting for ECOG, no significant difference was observed between treatments. The last three scenarios reflected the opposite clinical context, where more patients with ECOG > 1 received treatment A, but with no significant difference between treatments in unadjusted analysis. Adjusting analysis for ECOG revealed that patients with treatment B experienced more AEs over time. Semi-parametric prevalence is a complementary version of the non-parametric one, considering each covariate in estimating the probability of experiencing an AE over time. Confounding factors can be highlighted and considered to model the evolution of AEs over time according to patient’s clinico-pathological characteristics.