- This study aimed to find a simple, accurate way to determine key predictors of whether an individual patient will become pregnant after the first cycle of in vitro fertilization (IVF)
- The eIVF practice highway database was used to identify 22,413 patients who had their first IVF cycle with autologous oocytes and a plan for fresh embryo transfer
- Six machine learning models were tested that considered both conventional factors (e.g., demographics, infertility diagnosis) and, for the first time, in-cycle factors such as the number of oocytes retrieved
- The top statistically significant predictive variables were age (including the age categories of age > 42, age 41 to 42, age 38 to 40), number of cryopreserved embryos, and number of transferred embryos
- All models had an area under the curve (AUC) of about 0.70 for predicting pregnancy after the first IVF cycle, and one of them maintained high accuracy (AUC of 0.65) when only the top variables noted above were considered
Pre-cycle characteristics such as age, body mass index, infertility diagnosis, and reproductive history are widely used to predict the success of in vitro fertilization (IVF). Now, physicians at Massachusetts General Hospital in collaboration with Information and Systems Engineering Scientists at Boston University have created an in-cycle computer algorithm that also incorporates in-cycle parameters to predict whether an ongoing IVF cycle will result in pregnancy.
Subscribe to the latest updates from OB/GYN Advances in Motion
Karissa C. Hammer, MD, a clinical fellow in the Vincent Center for Reproductive Biology, Shruthi Mahalingaiah, MD, MS, a reproductive endocrinologist in the Department of Obstetrics & Gynecology, and colleagues describe the merits of the new prediction model in Scientific Reports.
From the eIVF practice highway database the researchers identified 22,413 patients who had their first IVF cycle between 2001 and 2018 with autologous oocytes and a plan for at least one fresh embryo transfer.
The researchers gathered data from multiple IVF centers in three different states (California, Massachusetts and New York) because state insurance mandates can influence how infertility is managed. Massachusetts is the only one of these states that mandates coverage for IVF procedures.
The team used 80% of the data to train six relatively simple machine learning models and tested them with the other 20% of the data. Machine learning is a form of artificial intelligence that self-improves from experience with large datasets without explicit programming. These computer algorithms provide more accurate and personalized prediction than regular statistical methods.
The variables considered in the models were:
- Pre-cycle factors—demographic and lifestyle factors, results of pre-cycle fertility evaluations and infertility diagnosis
- In-cycle factors—maximum estradiol value, number of oocytes retrieved, number of transferred embryos and number of embryos cryopreserved
Most Predictive Variables
The top five statistically significant variables that predicted pregnancy were:
- Age >42 (logistic regression coefficient, −0.31)
- Number of cryopreserved embryos (0.26)
- Age 41 to 42 (−0.25)
- Age 38 to 40 (−0.17)
- Number of transferred embryos (0.16)
Peak estradiol and the number of oocytes retrieved were the six and twelfth most important predictors, respectively.
Performance of the Prediction Models
XGBoost, an ensemble of decision tree models, achieved the best performance of the six models. The average area under the curve of the receiver operating characteristic (AUC) was 0.678. The other models performed very similarly with AUCs of 0.663 to 0.675.
In further analysis, the models were limited to considering only the top five most predictive variables. One of them, a regularized logistic regression algorithm, maintained high accuracy with an AUC of 0.65.
Advantages of the New Model
None of the models are currently intended for clinical use. However, a five-variable model would be easy to calculate and interpret in routine practice, unlike Bayesian networks and neural networks, for example.
Accurate prediction can help patients and their partners prepare psychologically for a positive or negative outcome while the first IVF cycle is underway. It also provides useful information to their physicians.
Predicting the likelihood of success after the first attempt can be especially helpful for counseling patients who live in states without mandated insurance coverage for IVF, because they may be able to afford only one treatment cycle.
Learn more about fertility care at Mass General
Refer a patient to the Mass General Fertility Center