Page 53 - The diagnostic work-up of women with postmenopausal bleeding

which are thought to be associated with an increased risk of endometrial cancer.
These variables have been tested in the original articles for univariate association and,
if sufficiently contributing to predictive accuracy in multivariable regression analysis,
combined to construct a clinical prediction model.We collected all different predictor
variables from the original articles, together with their significance, to identify the
most important predictor variables for endometrial cancer.The most important
predictor variables had been considered as statistically significant input variables in
three or more studies or were considered statistically significant in two studies and 3 had not been tested in other studies.
Model development assessment
The development of a prediction model consists of three phases: model derivation, model validation and impact analysis.25 In the first phase, model derivation, predictor variables are identified by logistic regression. Model validation, the second phase, consists of an internal and external validation phase.24 In internally validated models, the performance of the model is tested in the same data set in which the model was developed, or in a group of subsequent patients within the same centre. In external validation, the goal is to demonstrate generalizability and reproducibility in patients different from the patients used for derivation of the original model.Therefore, the prediction model is evaluated on new data collected from an appropriate patient population in a different centre.26 The final phase of model development is impact analysis, in which prediction models are tested for their ability to change clinician’s decisions and to change patient outcomes.27 All prediction models identified in this review are classified into the different phases of model development.We sent an email to all authors of the identified articles to investigate if their models are undergoing external validation and are not published yet.
Model performance
Performance measures (calibration, discrimination and clinical usefulness) and the range of probabilities given by the different prediction models were recorded. Calibration refers to the agreement between observed probabilities and predicted probabilities for groups of patients; this is usually reported as a calibration plot or a Hosmer-Lemeshow statistic (test for ‘goodness-of-fit’).28 Discrimination is commonly reported as the c-statistic (concordance), also referred to as the Area Under the receiver-operating characteristic Curve (AUC). It measures the ability of a prediction model in separating patients with endometrial cancer and patients without
Prediction models
51