Abstract
Clinical prediction models are increasingly being developed to support decision making in healthcare. For kidney transplantation, clinicians must quickly decide whether or not to accept a donor kidney for their patients based on a complex set of information about the donor, recipient and transplant process. A clinical prediction model could support their decision-making process. Following transplantation the survival process consists of semi-competing events, where one event censors another, but not vice versa. In kidney transplantation, death precludes graft failure, but graft failure does not preclude death. It is important to consider these aspects of the survival process when developing prediction models in the presence of semi-competing events.We first systematically reviewed all clinical prediction models for kidney transplant survival outcomes, and found that existing prediction models are poorly reported and most are considered to be a high risk of bias. In this review we identified a commonly used model, the Kidney Donor Risk Index (KDRI), and externally validated it in the UK kidney transplant population. The KDRI was well calibrated in the UK kidney transplant population, but showed moderate discrimination.
We proposed methods for simulating survival data from a multi-state model, with a fixed proportion of competing events. Using these methods, we conducted simulation studies to determine the impact of the competing event proportion on the predictive performance of a developed model. The prediction models were developed both with and without a semi-competing risks framework. We found that models that accounted for competing events provided better estimates of the discrimination when the competing event proportion was 0.3 and greater. However, estimation of the calibration slope was the same regardless of whether the competing event was accounted for or not. We conducted another simulation study to explore whether a separate or joint modelling approach should be used when developing multi-state models for prediction. We discovered that there was no difference in the discrimination regardless of the modelling approach, but the separate models were more computationally efficient.
| Date of Award | 2024 |
|---|---|
| Original language | English |
| Supervisor | Yinghui Wei (Director of Studies (First Supervisor)), Andrew Connor (Other Supervisor) & Kimberly Tam (Other Supervisor) |