In-Cycle Outcome Prediction Classification for IVF Treatment
T. Xu, S. Malhalingaiah, I. Paschalidis
T. Xu, S. Malhalingaiah, I. Paschalidis
In a study done with Boston University School of Medicine, the National Science Foundation, and the Center for Information Systems and Engineering, researchers aimed to create a method that would assist in the prediction of the success rate of a patient’s first IVF cycle. This study looked at patients who were doing their first IVF cycle and using their own eggs, and the method was tested against their outcomes.
In this article we’re going to examine:
When patients are interested in starting an IVF cycle, they will first need to determine if they are capable of using their own eggs-- or if donor eggs are required. Most women prefer to use their own eggs if possible, but sometimes factors such as age or health issues can affect this decision. Many women who are over the age of 40, and who want to use their own eggs, are at times discouraged from doing so. This is due to the fact that as a woman ages, her eggs also age as well. There is less likelihood that the IVF will be successful using the eggs of an over-40 woman.
This is not to say that it cannot be done, it just less likely. IVF is a costly procedure and most patients want to have the best chance possible at a pregnancy, so if they are recommended to do so, will accept the use of donor eggs. However, fertility clinics are available to treat over-40 women with a diminished ovarian reserve (DOR), and who want to use their own eggs. This is possible through customized IVF treatment plans and techniques such as early egg retrieval. This occurs when immature eggs are removed and allowed to mature in the lab via in vitro maturation (IVM). This increases the number of embryos available for implantation and can increase the chance of a pregnancy.
This study examined 22,763 patients using the data from the eIVF database, and sought to find a way to predict the outcomes of these IVF patients during their first cycle of fertility treatment. The prediction accuracy metrics used were AUC and AUC-PR. First, diagnoses were grouped into seven classifications, and a one-hot coding was used for all possible features. The data was randomly divided into test and training sets and feature standardization was used. The performance criteria was measured with AUC=0.5 with random guessing and AUC-PR=percentage of positive samples with random guessing.
For each patient, a series of medical factors were considered such as age, race, lifestyle, and body mass index (BMI). In addition, there were conditions relating to the egg and sperm taken into account as well. For each egg this included the total embryos cryoed, transfer count, and if the embryos were frozen after day five. For the sperm, the volume, concentration, motility, and progression were taken into consideration. In addition, hormone values were accounted for, and the data from day three ovarian test values were taken for: estradiol, progesterone, luteinizing hormone, follicle stimulating hormone, endometrium thickness, antral follicle count, as well as the maximum amount of estradiol.
The results were broken into four categories, age, race, diagnosis, BMI, and the outcome of pregnant and not pregnant was examined against these factors. For women less than 35, 55% of the sample were estimated to be pregnant and 40% were not. For women ages 38-40, only 16% had an outcome of pregnancy, while 21% did not. The chance of pregnancy drops off sharply after 40 years of age according to the analysis, with only 2% of women predicted to be pregnant and 7% were not.
When the results are broken down according to race, the researchers found that White women were the most likely to become pregnant at 22%, but this is also due to the fact that this race category contained the most patients. Asian women were predicted to have the outcome of pregnant after the first cycle of IVF at 4%, Hispanic/Latina at 2%, and Black women at 1%.
The reason for infertility was also examined, and when female infertility was the diagnosis, 45% of patients were estimated to be pregnant, and 14% were not. When there was a male infertility problem, 11% of the patients ended up pregnant while 6% did not have the outcome of pregnancy. If the diagnosis was tubal problems, it was an almost even percentage, with 5% pregnant and 4% not pregnant.
Finally, when the results are broken down according to BMI, the patients who had a BMI of 18.5-24.9 were the most likely to be pregnant, with 42% of women. As BMIs increase, the percentage of women who became pregnant drops. Those with BMIs lower than 18.5, only 2% would end up with a pregnant outcome.
The model used for this analysis achieved an accuracy of 69.11% in AUC and 67.57% in AUC-PR, which means this model is useful to help predict the success rate of an IVF treatment. As this study shows, the success rate of the first cycle of IVF is dependent on many things, each of which can play a part in a pregnancy result.