Page 58 - The diagnostic work-up of women with postmenopausal bleeding

Chapter 3
probability of cancer and the observed proportion of patients with endometrial cancer are mentioned in Randelzhofer et al39 However, calibration is generally reported as a calibration plot. None of the studies reported on calibration in a calibration plot.Discrimination was studied in seven out of eight articles by calculating an AUC.The AUC varied from 0.66 to 0.92 for different prediction models, with the highest AUC for a model combining Doppler and grey-scale TVS.37 In all internally validated studies clinical usefulness is described, with the highest sensitivity and the lowest negative LR for a combined model with patient characteristics, grey-scale TVS and Doppler.38 The highest NPV found for a model was 0.996 for a model, which combined patient history, endometrial thickness and histology in a sequential strategy.36 The performance of the four models using only patient characteristics showed a high sensitivity or high NPV in two models36, 38 and a low LR for a negative outcome in one model.38 All three studies in which Doppler was studied as a predictor variable, reported this information to contribute to the prediction of endometrial cancer in women with PMB.35, 37, 38 Endometrial thickness was used as a variable in eight prediction models and seven found that incorporating endometrial thickness may improve diagnostic accuracy of a model.
Discussion
We systematically reviewed existing prediction models for endometrial cancer in women with PMB and to identify the most important predictor variables.We found nine studies reporting on the development of prediction models for endometrial cancer in women with PMB. Eight of these studies described at least one aspect of internal validation and until now, none of the prediction models have been externally validated.
The different predictor variables can roughly be divided into four subjects: patient characteristics, grey-scale ultrasound variables, Doppler ultrasound variables and hysteroscopy variables. Most prediction models used a combination of these subjects to predict the chance of endometrial cancer. We chose to limit our list of most important predictor variables to those, which had been considered as statistically significant input variables in three or more studies and to those, which were significant input variables in two studies and had not been tested in other studies. By doing this, we identified the most important variables, without missing possible important variables, which have not yet been extensively studied. Using
56