What measurements are needed to predict pregnancy outcome in pregnancies of unknown location: does measuring hCG suffice?
Objectives: We aimed at investigating what information is sufficient for making a good prediction of the outcome of pregnancies of unknown location (PUL).
Abstract: 856 PULs were investigated at St Georges Hospital (London). There were 460 failing PULs, 330 intra-uterine pregnancies (IUP), and 66 ectopic pregnancies. Mathematical diagnostic models were constructed (a) using only hCG information, (b) using hCG and progesterone information, and (c) using an optimal set of measurements based on variable selection. The method used was multicategory logistic regression. Models were trained on a training set and evaluated on a test set using the area under the receiver operating characteristic curve (AUC). This was repeated 100 times in each of which the data were randomly splitted into training and test sets. The 100 resulting test set AUCs were summarized by their median. Three AUCs were computed each time: one for each outcome. Model (c) used hCG and progesterone information, the level of vaginal bleeding, and age. The median test set AUCs to predict failing PUL were 0.982 (a), 0.987 (b), and 0.987 (c). To predict IUP, we obtained 0.979 (a), 0.983 (b), and 0.984 (c). To predict ectopic pregnancy, we obtained 0.884 (a), 0.916 (b), and 0.931 (c). Thus, detection of failing PUL and IUP was easy using hCG information. Predicting ectopic pregnancy was more difficult. Mainly progesterone information but also age and the level of vaginal bleeding improved the prediction. The use of more advanced methods (support vector machines, kernel logistic regression) yielded similar conclusions.