Pregnancy Models Give Birth To New Health Insights
Having a baby is a life-changing decision that often requires a great deal of time and energy to ensure a positive outcome. But the cost of assisted reproductive technologies like artificial insemination or in-vitro fertilization (IVF) and the emotional impacts of infertility can be a lot to bear. To try to improve the chances of having a baby, Hariri Institute Research Fellow Yannis Paschalidis and an interdisciplinary team of medical researchers, including Lauren Wise at the BU School of Public Health and Shruthi Mahalingaiah at the Harvard T.H. Chan School of Public Health, used machine learning to create models that can predict the success of IVF procedures and natural pregnancies.
Many things can prevent a woman from becoming pregnant or carrying to term. IVF is one possible solution for couples that have fertility issues or cannot conceive naturally. In this process, eggs are collected and fertilized with sperm in a lab setting. The fertile eggs are then transferred to a uterus where they are implanted along the lining, and a successful IVF will result in the woman becoming pregnant. Although this method is helpful, it is far from perfect and the average IVF success rate is less than 60%.
A lot of data is collected both before and during pregnancy to monitor the health of the mother and baby. If the mother goes in for regular health checks, things like her weight, diet, and habits are all recorded. But most current fertility models only focus on one health factor at a time. This is where the researchers saw great possibilities. Paschalidis is an expert in machine learning and he tries to apply such methodologies to as many disciplines as possible. “In general, I think there is a great opportunity to use data science and machine learning in public health,” he said, “We live in an era where there is a lot of data available.”
By using large databases of health records, the team trained machine learning algorithms to predict both the outcomes of natural pregnancies and IVF procedures. The findings were published recently in Nature Scientific Reports and Human Reproduction.
Estradiol levels predict pregnancy outcomes
The researchers looked at many different health factors in women who were trying to conceive. Using their models, the researchers found that certain health data are important for predicting the success of natural or IVF fertilization. The mother’s age, number of cryopreserved embryos, estradiol levels, body mass index (BMI), diet, menstrual cycle length, and stress level are all important to ensuring a positive outcome. Each factor was examined together rather than individually, unlike earlier models. And some data, like estradiol levels, were not considered previously at all.
These findings highlight the importance of testing hormone levels in women. Estradiol levels are not often collected because they are more difficult to gather than other health measurements like weight and age, according to Paschalidis. However, hormone levels can be important indicators of balance, or imbalances, in the body. They are explicitly linked to the menstrual cycle, brain function, metabolism, body growth, and more.
Racial differences in IVF success rates
Paschalidis and colleagues also found a new predictive factor for the outcome of IVF – race. The researchers’ model determined that white women have more successful pregnancy outcomes on average. This could be because the IVF process has been optimized for white women when it was first created and tested. The racial difference in success could also be due to underlying disparities such as overall health, income levels, and age when trying to conceive. One important caveat is that the majority of people participating in this study did not disclose their race. Only by working with a large and diverse dataset can the researchers quantify these systemic problems in the future. “Having more data from multiple groups would strengthen the model and also strengthen our confidence,” said Paschalidis.
Democratizing the use of predictive modeling
One of Paschalidis and colleagues’ models can predict IVF outcomes with 67% accuracy, and the other can predict the success of natural pregnancies with 70% accuracy. Both of these models are more accurate than current methods, and are promising for the future of AI and healthcare integration. Additionally, both of these models offer a more holistic approach to predicting pregnancy outcomes.
The accuracy of the models is due to the unique way that they were created. Normally, when researchers provide datasets to a model, so called outliers, or data points that are quite different from others, may fool the algorithm. Paschalidis and colleagues developed methods that could deal with such outliers in the learning process. “We trained these models to be in many ways robust to potential outliers,” said Paschalidis, “If you have outliers in model training, the machine learns to accommodate those, and becomes worse in regular predictions needed later on. Our robust learning methods avoid this issue.”
Models are one of the tools that can advance equity in the field of healthcare, and make sure that everyone has equal access to the resources that they want and need to thrive. “One goal that we have is to democratize the use of predictive modeling in IVF,” Paschalidis said. These procedures can be very expensive and are not always covered by insurance, thus introducing inequities into the healthcare system. With such a large financial commitment required, a free predictive model could help inform decisions by individuals in consultation with their physicians.
But Paschalidis notes that models will never replace clinicians who are able to evaluate other factors in real time when they are seeing a patient. “An algorithm just responds to the data given to it, and there may be other factors that are not captured,” said Paschalidis, “These models should be used by clinicians and patients as additional information or guidance but the people should be the ones making decisions.”
BY ALEX JOHNSON