Kidney transplant is the main remedy for end-stage renal disease and the prognosis of allograft survival is what recipients care about the most. A popular method for allograft survival prediction in kidney transplantation is through the Cox proportional hazard model. There is a substantial literature and the performance of the published models varies greatly. One possible explanation driving this variability of performance is the high heterogeneity that is intrinsic in the transplant population. We propose two complementary approaches (bottom-up and top-down) that aim to identify recipient sub cohorts based on the inherent structure of the data which will improve allograft survival. The innovations of our approaches lie in combining supervised and unsupervised learning, that is, it integrates machine learning methods with survival analysis. The bottom-up approach uses Numero, a new self-organising-map method, with the elastic net Cox model to stratify potential recipient sub cohorts. The alternative top-down approach uses the Cox model with a contrast tree method to identify cohort characteristics. Examining the results from both approaches, we find that recipient waiting time is an important predictor in predicting graft survival for the whole population. We also find that there is a large amount of heterogeneity among âunfitâ recipients, these recipients have sub cohorts that are particularly hard to predict in terms of their graft survival. In contrast, for younger and âfitâ cohorts, we found that immunological factors are important components. The ability to identify sub cohorts based on prediction outcome is useful for enhancing prediction of graft survival and has the potential guide allocation algorithm.