External validation of prognostic models: what, why, how, when and where?

Abstract

Prognostic models that aim to improve the prediction of clinical events, individualized treatment and decision-making are increasingly being developed and published. However, relatively few models are externally validated and validation by independent researchers is rare. External validation is necessary to determine a prediction model's reproducibility and generalizability to new and different patients. Various methodological considerations are important when assessing or designing an external validation study. In this article, an overview is provided of these considerations, starting with what external validation is, what types of external validation can be distinguished and why such studies are a crucial step towards the clinical implementation of accurate prediction models. Statistical analyses and interpretation of external validation results are reviewed in an intuitive manner and considerations for selecting an appropriate existing prediction model and external validation population are discussed. This study enables clinicians and researchers to gain a deeper understanding of how to interpret model validation results and how to translate these results to their own patient population.

Keywords: educational; external validation; methodology; prediction models.

© The Author(s) 2020. Published by Oxford University Press on behalf of ERA-EDTA.

Figures

FIGURE 1

Illustration of different validation types. A developed prediction model can be validated in various ways and in populations that differ from the development cohort to varying degrees. Internal validation uses the patients from the development population and can therefore always be performed. As internal validation does not include new patients, it mainly provides information on the reproducibility of the prediction model. Temporal validation is often considered to lie midway between internal and external validation. It entails validating the model on new patients who were included in the same study as patients from the development cohort but sampled at an earlier or later time point. It provides some information on both the reproducibility and generalizability of a model. External validation mainly provides evidence on the generalizability to various different patient populations. Patients included in external validation studies may differ from the development population in various ways: they may be from different countries (geographic validation), from different types of care facilities or have different general characteristics (e.g. frail older patients versus fit young patients). Not every model needs to be validated in all the ways depicted. In certain cases, internal validation or only geographic external validation may be sufficient; this is dependent on the research question and size of the development cohort.

FIGURE 2

Cumulative histogram of the number of hits on PubMed when using a simple search strategy of prediction models and adding external validation to this search. Search strategies are given in Appendix A. PubMed was searched from 1961 to 2019. The total number of prediction model studies retrieved was 84 032, of which 4309 were found when adding an external validation search term. The percentage of studies with external validation increased over the years; in 1990, 0.5% of published prediction studies mentioned external validation, while in 2019 this was 7%.

FIGURE 3

Example of a calibration plot. The dotted line at 45 degrees indicates perfect calibration, as predicted and observed probabilities are equal. The 10 dots represent tenths of the population divided based on predicted probability. The 10% of patients with the lowest predicted probability are grouped together. Within this group the average predicted risk and proportion of patients who experience the outcome (observed probability) are computed. This is repeated for subsequent tenths of the patient population. The blue line is a smoothed lowess line. For a logistic model this is computed by plotting each patient individually according to their predicted probability and outcome (0 or 1) and plotting a flexible averaged line based on these points. In this example calibration plot we can see that the model overpredicts risk; when the predicted risk is 60%, the observed risk is ∼35%. This overprediction is more extreme for the high-risk x-axis. If a prediction model has suggested cut-off points for risk groups, then we recommend plotting these various risk groups in the calibration plot (instead of tenths of the population).