Prospective validation of a clinical prediction model for diagnosing adenomyosis with ultrasound

Ultrasound diagnosis of adenomyosis involves many diagnostic signs that vary in their diagnostic significance. Our objective was to validate a previously developed multivariate prediction model for diagnosing adenomyosis that weights predictors based on their diagnostic performance and evaluate the model's performance between less experienced examiners and experts. Prospective, observational, multicentre study including premenopausal women scheduled for hysterectomy for benign indications. All patients underwent preoperative 2D and 3D-transvaginal ultrasound. Nine diagnostic predictors as previously defined (hyperechoic islets, fan-shaped shadowing, myometrial cysts, globular or normal uterine shape, uterine wall thickest/thinnest ratio, maximum width of the junctional zone (JZ) in sagittal plane, regular JZ, grade of dysmenorrhea) were documented. Images, videos and 3D-volumes were stored and re-assessed by a blinded expert. Histopathological examination of the hysterectomy specimen followed a dedicated protocol. The weight of each sign (β) was predefined earlier; the diagnostic performance [area under the curve (AUC), sensitivity, specificity] of the model was evaluated. Of 528 women, 465 were included in the final analysis. Adenomyosis was present in 255/465 (54.8%). The diagnostic performance of the model in the hand of non-experts showed a poor test quality [AUC= 0.61 (95%CI 0.55-0.66), sensitivity 65%, specificity 58%]. When asked to subjectively assess the presence of adenomyosis, the overall performance was comparably low [sensitivity 72% (95%CI 0.66-0.77), specificity 54% (95%CI 0.47-0.61), PPV 65%, NPV 61%, accuracy 64%]. The results for the test quality for the expert reader are pending and will be presented during the congress. The use of hormonal medication did not influence the model's performance. The preliminary results of the validated model show a poor test quality for non-expert readers. Adjustment of the predictors might improve diagnostic quality.