In reproductive medicine, we have a wealth of data with which to make personalized predictions of treatment outcomes for patients. The types of clinical predictors (or variables) aren’t new, but only recently with machine learning and cloud computing technologies have we been able to use it to help fertility patients in an impactful way.

My research team and I worked hard to build a validated IVF prediction model because we knew that more women would realize their dreams of building a family if only they knew their personalized probability of IVF success. As a doctor, it’s painful to see patients who have waited too long to seek help, simply because they weren’t aware just how much a delay in treatment can impact their IVF success rates. It is also equally painful to see women with very good chances of IVF success quit treatment after their first IVF cycle fails, due to the financial risk of unsuccessful treatment. On the flip side, millions of American couples experiencing infertility wrongly assume there is no hope for them. There is a lot we can do to educate them about how IVF can help them have a family.

AI/machine learning has the power to address these circumstances by helping women learn about the importance of seeking help sooner rather than later, and providing more precise predictions of success for women who might otherwise assume their chances of having a baby are slim.

Complex Data and Machine Learning

While age is a significant predictor of a woman’s chances of having a baby from IVF, we have found that it consistently contributes to only 50% of the prediction. That means it is important to use other predictors which make up the other 50% of the prediction, even though each of those predictors on its own may contribute to only 5 to 10% of the prediction.

Markers of ovarian reserve, embryo quality measures, reproductive history, and health data such as body mass index (BMI) and chronological age have all been individually associated with fertility treatment outcomes. However, unless these factors are built into a prediction model that can be objectively and quantitatively tested, the information they carry is not easily and accurately accessed because there is a tremendous amount of correlation and redundancy of information among these predictors.

For example, if a woman is “older,” at 41, and her AMH level is 1.8 (good for 41), how does her IVF success probability compare with a woman who is “younger”, at 36, whose AMH level is 0.8?  How do you advise each of these women? Using machine learning and our proprietary programs developed from more than a decade of analyzing IVF outcome data of a diverse and global group of collaborating IVF centers, the Univfy data science team can compile the odds of increased or decreased live birth rates of each clinical factor into one composite number that would show the probability of having a baby in a way that is meaningful to a fertility patient.

We have found that machine learning is a very powerful way to tackle complex data sets encountered in the IVF space. We're already using machine learning in our everyday lives: It runs most internet search engines, recommendations from Amazon and Netflix, and the "people you may know" feature on Facebook or Linkedin. In other words, we are afforded more accuracy and personalization in our online shopping experiences than most patients receive when making emotional, life-affirming and costly medical decisions such as moving forward with IVF treatment to have a family.

Discarding Assumptions

To apply machine learning to health data and predict IVF outcomes, we first had to peel away some long-held assumptions in fertility statistics. We had to assume that we did not know which of many clinical factors, such as age, BMI, clinical diagnosis, or sperm count are strong outcome predictors, and how they rank in terms of predictive value. By doing this reset, we were able to open the way to answering many complex questions.

About the Author:

Mylene Yao, M.D. | Co-Founder and CEO of Univfy®

Dr. Mylene Yao is a board-certified OB/GYN with more than 20 years of experience in clinical and reproductive medicine research. Prior to founding Univfy, she was on the faculty at Stanford University, where she led NIH-funded fertility and embryo genetics research and developed The Univfy AI Platform with the academic founding team. See her full bio here.