Development and validation of prediction models for endometrial cancer in postmenopausal bleeding

Objective To develop and assess the accuracy of risk prediction models to diagnose endometrial cancer in women having postmenopausal bleeding (PMB). Methods A retrospective cohort study of 4383 women in a One-stop PMB clinic from a university teaching hospital in Hong Kong. Clinical risk factors, transvaginal ultrasonic measurement of endometrial thickness (ET) and endometrial histology were obtained from consecutive women between 2002 and 2013. Two models to predict risk of endometrial cancer were developed and assessed, one based on patient characteristics alone and a second incorporated ET with patient characteristics. Endometrial histology was used as the reference standard. The split-sample internal validation and bootstrapping technique were adopted. The optimal threshold for prediction of endometrial cancer by the final models was determined using a receiver-operating characteristics (ROC) curve and Youden Index. The diagnostic gain was compared to a reference strategy of measuring ET only by comparing the AUC using the Delong test. Results Out of 4383 women with PMB, 168 (3.8%) were diagnosed with endometrial cancer. ET alone had an area under curve (AUC) of 0.92 (95% confidence intervals [CIs] 0.89-0.94). In the patient characteristics only model, independent predictors of cancer were age at presentation, age at menopause, body mass index, nulliparity and recurrent vaginal bleeding. The AUC and Youdens Index of the patient characteristic only model were respectively 0.73 (95% CI 0.67-0.80) and 0.72 (Sensitivity = 66.5%; Specificity = 68.9%; +ve LR = 2.14; -ve LR = 0.49). ET, age at presentation, nulliparity and recurrent vaginal bleeding were independent predictors in the patient characteristics plus ET model. The AUC and Youdens Index of the patient characteristic plus ET model where respectively 0.92 (95% CI 0.88-0.96) and 0.71 (Sensitivity = 82.7%; Specificity = 88.3%; +ve LR = 6.38; -ve LR = 0.2). Comparison of AUC indicated that a history alone model was inferior to a model using ET alone (difference = 0.19, 95% CI 0.15-0.24; p < 0.0001) and History plus ET (difference = 0.19, 95% CI 0.16-0.23, p < 0.0001) and history plus ET was similar to that of using ET alone (difference = 0.001 95% CI -0.015 to 0.0018, p = 0.84). Conclusions A risk model using only patient characteristics showed fair diagnostic accuracy. Addition of patient characteristics to ET did not improve the diagnostic accuracy as compared to ET alone in our cohort.