External Validation
The reliability and acceptability of a prediction model largely depend
on how well it performs in a validation cohort, outside of the
derivation cohort where the model was developed. Internal validation of
prediction models is often not sufficient for generalizability, and
external validation is necessary before implementing prediction models
in clinical practice. External validation of models is often considered
essential to support the general applicability of a prediction model as
it addresses transportability1, 2. Transportability
requires the model to perform accurately in predicting data drawn from a
different but plausibly related population or in data collected by using
a little different method than those used in the development
sample3. External validation requires data collected
from a similar group of patients in a different setting and aims to
address the accuracy and performance of a prediction model in a
different patient population. These data (sample) are fully independent
of the development data and originate from different but similar
patients. The generalizability of a model becomes stronger when the
model is externally validated multiple times and in a more diverse
setting2. This is the reason perhaps why the
Framingham Risk Score (FRS) for cardiovascular disease (CVD) is so
widely used in the clinical setting as the model was externally
validated many times with many different settings.
Most studies evaluating prediction models focus on the issue of internal
validity as opposed to the important issue of external validity.
Internal validation does not guarantee generalizability, and thus
external validation is necessary before implementing prediction models
into clinical practice.
External validation can be assessed in different ways:
- Temporal validation (validation in more recent individuals)
- Geographical validation (validation in other places)
- Fully independent validation/Strong external validation (by other
investigators at other sites)