The first prospective evaluation of the IOTA phase 1 mathematical models to predict the character of an adnexal mass

Objectives: During IOTA phase I, more than 50 sonographic and demographic variables of 1066 patients with an adnexal mass were collected in nine European centers. All patients were scanned by a strict protocol. This database was used to develop several mathematical models that could be useful for the preoperative prediction of the character of an adnexal mass. The aim of IOTA phase Ib was to collect a new dataset for prospective testing of all of the previously developed models. Methods: Prospective model performance was assessed from the area under the ROC curve (AUC), accuracy, sensitivity, specificity, PPV and NPV. Results: Some 507 patients from three IOTA centers were included; 28% (142) masses were malignant. The AUC of the logistic regression model was 0.942 on the test set of IOTA 1 (n = 312), 0.950 on the new dataset, and 0.954, 0.908 and 0.992 in each of the individual centers. For the least squares support vector machine model with linear kernel the AUC on the IOTA 1 test set was 0.946, 0.950 on the new dataset, and 0.950, 0.894 and 0.986 for each center individually. For the relevance vector machine with linear kernel the AUC was 0.949 on the test set of IOTA 1, 0.947 on the new dataset, and 0.947, 0.896 and 0.984 for each center individually. A multilayer perceptron neural network, finally, had an AUC of 0.942 on the test set of IOTA 1, 0.948 on the new dataset, and 0.947, 0.917 and 0.985 for each center individually. Conclusions: The analysis shows that all models performed well when tested prospectively on the whole dataset and in each center individually. The next step will be to test the models prospectively in new centers with different levels of ultrasound experience and population characteristics.