Page 85 - The diagnostic work-up of women with postmenopausal bleeding

External validation of a prediction model
Discussion
In this external validation study, we demonstrate that the diagnostic performance (i.e. Discrimination and calibration) in external validation of the two models is similar to the discriminative performance of the models in internal validation in the development
database.16 The model with the best discriminative performance in this external
validation is the patient ‘characteristics and TVS’ model. However, we found that the discriminative performance of the ‘patient characteristics and TVS’ model is similar to
that of endometrial thickness measurements with TVS (TVS-only), i.e. adding clinical
information to endometrial thickness measurement does not significantly improve
the ability to discriminate between benign and malignant endometrium. 4
Applying a strategy basing the decision to perform further invasive diagnostics on an individual risk calculated with ‘patient characteristics only’ would be safe.This means, no endometrial (pre) cancer that would have been detected by selecting women with TVS would be missed if these women would be selected based on patient characteristics only. However, you would need to perform invasive diagnostics in 93% of women, compared to only 61-63% (respectively in the Dutch and Swedish database) when patients would be selected based on TVS.
An important strength of our study is the external validation of the models using data from a different region within the Netherlands as well as data from another European country. External validation, assessing the validity and generalizability of a model is an essential step before a model can be implemented in practice.21,23 To our knowledge this is the first study to describe external validation of a prediction model estimating the risk of endometrial cancer in women with PMB. As the TweeSteden Hospital also participated in the development study of the two prediction models by Opmeer et al, the population used for external validation has a minor overlap with the development population, yet with completely separate samples (different women in the development and validation sample).
As many data were collected as part of clinical practice, not all information was available for all women. Multiple imputation was used to deal with these missing data. Multiple imputation, even with a relatively large amount of missing data, gives a more precise and valid measure of association for variables with missing values than complete case analysis.17,21 Generally, dropping cases with missing values (complete case analysis) yields biased results, and the discriminative ability of a multivariable model is reduced when cases with missing values are excluded from analysis.21
83