Page 57 - The diagnostic work-up of women with postmenopausal bleeding

Predictor variables
The nine included articles investigated 27 different possible prediction variables
(Table 2).Age was tested in all nine articles,turned out to be statistically significant
in multivariable analysis in six articles and was used in the prediction model in six
articles. Endometrial thickness was tested in eight articles, statistically significant in
multivariable analysis in eight articles and used in eight prediction models. Most
important predictor variables in patient history were: age, body mass index (BMI),
diabetes, frequency of bleeding, use of anticoagulants and HRT. Endometrial thickness, 3 endometrial morphology and endometrial border were identified as significant grey-
scale TVS variables. In the three articles studying the use of Doppler for predicting endometrial cancer, endometrial colour score and vascularity index were identified as the most important predictor variables.
Phases of model development
All articles selected in this review addressed the first phase of developing a prediction model: model derivation.24 Of the nine articles on predicting endometrial cancer in women with PMB,eight had been internally validated but none of these models passed the external validation phase.We asked all six research-groups, which developed the nine different prediction models if their models are undergoing external validation and we received response from all six research-groups.The two prediction models of Opolskiene et al,37, 38 are undergoing temporal validation (internal validation in a newly recruited patient group) and external validation in an international multicentre study by Valentin et al No results are available yet, since they are still recruiting patients for these studies.The two prediction models developed by Burbos et al33, 34 were recently used in an article to compare the performance in internal validation of these models.43 This group is working on external validation. Finally, we can report that the prediction model of Opmeer et al36 is currently being externally validated in two cohorts: one cohort in three different hospitals in the Netherlands and one in Skåne University Hospital Malmö in collaboration with the group of Valentin et al, but this external validation is not published yet. There were no impact analysis studies, i.e. studies that showed that the prediction model indeed improved patient outcome or was cost-effective in clinical practice.
Performance of the prediction models
The performance of the eight articles that were internally validated their models is presented inTable 3.31, 33-39 Calibration was described in one article.39 The estimated
Prediction models
55