Investigation of the performance of mathematical models on small ovarian masses in the IOTA phase 1 and 2 study

The first phase of IOTA resulted in a data set of 1066 patients from 9 centers in 5 countries. Previously, this data set was randomly stratified in 70% of patient data to construct a logistic regression model (referred to as M1) and 30% of the patient data as a test set to estimate the predictive performance. IOTA phase 2 resulted in a data set of 1940 patients from 19 centers in 8 countries. We investigate whether the performance of M1 depends on the size of ovarian masses when used prospectively on the IOTA phase 2 data set. Jump to… Methods The performance of M1 was estimated on the IOTA phase 2 data set by calculating the Area Under the ROC curve (AUC) on patients with a maximum lesion diameter smaller than a predefined threshold. This threshold was then iteratively increased to investigate the evolution of the AUC as a function of the size of the ovarian mass. Jump to… Results We observed a significant decrease of the AUC on the IOTA phase 2 data set when the maximum diameter of the lesion is increased from 28 mm to 32 mm. The AUC for all masses with a maximum lesion diameter smaller than 29 mm is 0.947 (SE 0.023) while the AUC for all masses with a maximum lesion diameter smaller than 33 mm is 0.889 (SE 0.047). When focusing on this subgroup of patients with ovarian masses with a maximum lesion diameter from 29 to 32 mm, we found 61 patients which were significantly younger (P-value = 0.057) and had a different color score distribution (P-value 0.0066) compared to the remaining patients from the IOTA 2 data set. There were 6 malignant masses (10%) in this set of patients, while M1 predicted 14 masses to be malignant (4 correct) when using the previously determined threshold of 0.1 for classifying ovarian masses. Similar results were observed on the smaller IOTA phase 1 test set. Jump to… Conclusions These results indicate that masses with a maximum lesion diameter from 29 till 32 mm are hard to classify for mathematical models.