PLoS ONE (Jan 2023)

Predicting mortality after start of long-term dialysis–International validation of one- and two-year prediction models

  • Mikko Haapio,
  • Merel van Diepen,
  • Retha Steenkamp,
  • Jaakko Helve,
  • Friedo W. Dekker,
  • Fergus Caskey,
  • Patrik Finne

Journal volume & issue
Vol. 18, no. 2

Abstract

Read online

Background Mortality prediction is critical on long-term kidney replacement therapy (KRT), both for individual treatment decisions and resource planning. Many mortality prediction models already exist, but as a major shortcoming most of them have only been validated internally. This leaves reliability and usefulness of these models in other KRT populations, especially foreign, unknown. Previously two models were constructed for one- and two-year mortality prediction of Finnish patients starting long-term dialysis. These models are here internationally validated in KRT populations of the Dutch NECOSAD Study and the UK Renal Registry (UKRR). Methods We validated the models externally on 2051 NECOSAD patients and on two UKRR patient cohorts (5328 and 45493 patients). We performed multiple imputation for missing data, used c-statistic (AUC) to assess discrimination, and evaluated calibration by plotting average estimated probability of death against observed risk of death. Results Both prediction models performed well in the NECOSAD population (AUC 0.79 for the one-year model and 0.78 for the two-year model). In the UKRR populations, performance was slightly weaker (AUCs: 0.73 and 0.74). These are to be compared to the earlier external validation in a Finnish cohort (AUCs: 0.77 and 0.74). In all tested populations, our models performed better for PD than HD patients. Level of death risk (i.e., calibration) was well estimated by the one-year model in all cohorts but was somewhat overestimated by the two-year model. Conclusions Our prediction models showed good performance not only in the Finnish but in foreign KRT populations as well. Compared to the other existing models, the current models have equal or better performance and fewer variables, thus increasing models’ usability. The models are easily accessible on the web. These results encourage implementing the models into clinical decision-making widely among European KRT populations.