Page 59 - The diagnostic work-up of women with postmenopausal bleeding

these limits we identified 11 important input variables for predicting endometrial cancer in women with PMB (Table 2).
Almost all articles reported performance in terms of discrimination and/or clinical usefulness, whereas calibration was reported only incidentally. In this study, we identified five articles describing a prediction model with good discrimination (AUC of >0.8).31, 35-38 Because only one study described data on calibration, there is insufficient data available to draw conclusions on calibration.
Two studies showed best performance regarding discrimination and clinical 3 usefulness: Opolskiene et al 2011 and Opmeer et al 2007. In the model by Opolskiene
et al 2011, a combination of patient characteristics, grey-scale TVS and Doppler was
used.They concluded that their model excludes endometrial cancer reasonably well
when power Doppler is added. Furthermore, in all three studies that used Doppler,
Doppler was found to contribute to the prediction of endometrial cancer in women
with PMB.35, 37, 38 Based on this, we could conclude that the best model in predicting
endometrial cancer is a model, which uses a combination of patient characteristics,
endometrial thickness and power Doppler. However, power Doppler cannot be
used in all patients. All three Doppler-models excluded patients based on different reasons:Doppler artefacts,incorrect processing ofTVS image,fluid in the cavity and
absence of Doppler signals or large myomas.Another limitation in the use of power
Doppler is that these studies do not give information on the interobserver variability
and learning curve in measuring Doppler variables. For application of results found
in Doppler studies, it is important to use the same ultrasound system, as the colour
content of a power Doppler scan depends heavily on Doppler sensitivity.38
Although the performance of the models using Doppler seems reasonable, a model using patient characteristics and endometrial thickness might be more useful in daily practice. In a health care system with general practitioners referring patients with a high risk of malignant disease to a specialist, the best model would be a model that can distinguish women with a high risk of endometrial cancer from women with a low risk based on patient characteristics only. Such a model would also be useful in situations where TVS is not directly available. Only women with a high risk could be referred forTVS or to the gynaecologist for a further evaluation and women with a low risk could be reassured and referred only at recurrent bleeding. Based on this review we couldn’t identify a model with a good performance in internal validation based on patient characteristics only. However, two of four models based on patient characteristics showed good performance in clinical usefulness with a high sensitivity, a high NPV and/or a low LR for a negative outcome.36,38 Based on these results we
Prediction models
57