Developing a better crystal ball to predict IVF success: IVF prediction softwareNovember 2, 2012Carole 2 Comments »
Before we start down the road to IVF, is there a way to determine the probability that IVF treatments will work for us? Is there a way to predict our risk of multiples when we are being asked to decide between transferring one or two…or even three embryos? I failed my first IVF, what are my chances that a second cycle would work for me?
Patients ask themselves and their doctors these questions all the time and doctors estimate chances for IVF success and multiple rate based on the average experience of thousands of patients who share the same age or other factors and also their own first-hand experience with patients in their practice. When predicting IVF success, maternal age is given the most weight, based on thousands of cycles reported to the CDC each year. However, an individual patient’s specific probability of success is affected by other factors besides age such as body mass index, patient diagnosis, FSH level on day 3 and semen analysis results.
For you statistics buffs out there, the analysis is based on something called “boosted tree” regression analysis. Copied from their website is this explanation: “Boosted tree, or boosting, is a subtype of the regression tree, a statistical method used to develop prediction models used in many fields. Univfy’s prediction model uses the answers to a specific series of questions to determine which set of questions to “branch off to” next. In this case, one set of questions determines the next set of questions to ask, much like how a branch of a tree leads to another unique set of branches until the analysis is complete.” There’s more statistical explanation but the point is that this type of “boosted tree” analysis amplifies its predictive power by identifying more and less important factors based on cumulative responses to patient specific questions.
Specifically, the Univfy software evaluates various patient factors including age, body mass index, fertility diagnoses, response to hormonal treatment, embryo quality scores, the amount of hormonal drugs used, patient age, and the thickness of the uterine lining on ultrasound, number of embryos and spouse’s age and comes up with the probability of specific outcomes (pregnancy or multiples) given that data set. Because it looks for relationships between patient specific data and outcomes, it can use indicators of success, like high estradiol level during the cycle without having to understand the molecular pathway underlying high estradiol level in the IVF cycle and successful pregnancy. The video, “Univfy’s Approach” , explains their statistical prediction model.
If you want to read their original research papers, you can find them here (copied from the website):
Jun SH, Choi B, Shahine L, Westphal LM, Behr B, Reijo Pera RA, Wong WH, Yao MWM. Defining human embryo phenotypes by cohort-specific prognostic factors. PLoS ONE 2008;3(7): e2562. doi:10.1371/journal.pone.0002562. The full-length article is available at no charge on the PLoS ONE web site: http://www.plosone.org/article/info%3Adoi%2F10.1371%2Fjournal.pone.0002562
Banerjee P, Choi B, Shahine LK, Jun SH, O’Leary K, Lathi RB, Westphal LM, Wong WH, Yao MWM. Deep phenotyping to predict live birth outcomes in in vitro fertilization. PNAS 2010;107(31):13559-60. The full-length article is available at no charge on the PNAS web site: http://www.pnas.org/content/107/31/13570.long
Lannon BM, Choi B, Hacker MR, Dodge LE, Malizia BA, Barrett CB, Wong WH, Yao MWM, Penzias AS. Predicting personalized multiple birth risks after in vitro fertilization – double embryo transfer. Fertil Steril 2012. doi:10.1016/j.fertnstert.2012.0411. The abstract of this research article can be viewed at the Fertility and Sterility website: http://www.fertstert.org/article/S0015-0282(12)00443-8/abstract
What is intriguing about this software is its potential to give patients a crystal ball of sorts, which could empower patients to make better decisions about whether to pursue IVF at all (PreIVF), stop after one cycle (PredictIVF) or reduce the number of embryos transferred per cycle (IVF Single).
Which patients can benefit from prediction models to make IVF-related decsions? Too many times, patients who have little chance of becoming pregnant with IVF continue to endure cycle after cycle, wasting both their time and money when they could be pursuing other options for family building or making other choices. Another scenario is the patient who really has a pretty good chance of success but gives up after one failed cycle –to drop out completely, switch clinics or delay making any decision–all of which can make pregnancy less likely. Unfortunately, there is often a learning curve for doctors with every patient. Some patients don’t respond as expected to medications and the treatment plan for the next cycle is often adjusted based on the experience of the first cycle. Switching clinics means starting over at zero and hoping the next doctor will learn from the first failed cycle, rather than just make similar mistakes with a “first cycle” at the second clinic. PredictIVF is the test that uses additional information gleaned from the first failed IVF test to predict how likely it is that a second cycle will work.
Predicting the chance of pregnancy from the first cycle of IVF. PreIVF is offered to patients who have never used IVF before to see what their chance of success with IVF is likely to be in a first cycle. I interviewed co-founder Dr. Mylene Yao for this post. One of the questions I had was what tests or diagnostic work-up, if any, did the patient need to use PreIVF because the test is applied before the first diagnostic visit. She explained that the only two tests required are a semen analysis and day 3 FSH blood level; both of these can be ordered by a family doctor or ObGyn.
The test costs $250.00 (or about the cost of the first patient visit with an REI) but is not covered by insurance. Interestingly, Univfy has eligibility requirements patients need to meet before they can buy a test. This actually benefits the patient because the eligibility requirements rule out patients who have situations or conditions that were not considered in building the prediction model so the test will not be valid for these patients at this time, though efforts are underway to build better models for these patients. For instance, if you have the following situation (copied from the website) , PreIVF is not a valid test for you:
- You have had IVF treatment in the past;
- You are 43 years old or above;
- Your serum Day 3 FSH level is greater than 14 U/L;
- Your antral follicle count (AFC) is less than 6;
- You are planning to do IVF with donor eggs or a gestational carrier;
- You are planning to do IVF for fertility preservation or pre-implantation genetic diagnosis for hereditary disease, rather than for infertility.
Dr. Yao readily acknowledges that the prediction model can not consider every possible factor (particularly if factors are non-reproductive ones like other chronic medical conditions; diabetes, multiple sclerosis, high blood pressure etc,) which would modify a physician’s recommendation regarding IVF as a treatment plan. The power of this software is that it can highlight important factors that patients can then bring up with their doctor and also provide an opening for the doctor to discuss other factors- clinic specific or patient specific -that may also affect the patient’s chances of success.
Prediction software is better able to consider many more predictive factors at once than humans can. Some doctors will make their recommendation regarding the number of embryos to transfer based on patient age and transfer the recommended number for that age as per SART guidelines without much variation in practice for all patients. Often the quality of embryos is also taken into account. A more sophisticated approach is to consider even more factors such as the patient’s prior history with IVF and pregnancy, what the cohort of embryos look like and even hormonal factors such as peak serum E2. The predictive model is able to take more patient specific factors into account, providing a more individual prediction of the risk of multiples than the simplest predictions based on maternal age only.
Now it’s important to remember that no prediction model is perfect and so should be considered as only one factor when making these important decisions with your doctor. Not all the tests are currently available direct to patient customers. Predict Single is currently offered only as a customized clinic test–meaning the clinic has to enroll and provide data to Univfy– because specific clinic factors also play a role in a specific patient’s risk of multiples. Clinics have the option to participate with Univfy to create a customized prediction model using clinic- specific data. For these clinics, the predictive model can distinguish between patients who are in the low, medium or high success group within the their clinic’s patient population.
Which brings up another point, the model can only predict your probability of success at a top quality clinic. You might have very few patient-specific barriers to IVF success but a poor quality clinic can significantly reduce or eliminate your chances of pregnancy. In past years the clinic pregnancy rates reported to the CDC in the youngest age group ranged from 9% to over 70%–and that range certainly has more to do with lab and physician quality than patient cherry picking. So it is still up to the patient to find the best clinic they can get to.
What I like about this predictive model approach is that it gives patients more information and a good starting point for an in-depth personal discussion with their doctor. Ironically, using prediction modeling software that can look at lots of variables may enable patients to get more customized care as individuals and avoid feeling like just a number in the clinic with a one size fits all treatment plan. My hope is that patients can arm themselves with more knowledge to ask better, specific questions about their treatment plan. More tools, more information should lead to better patient discussions with their healthcare team, better treatment plan decisions for the patient and ultimately, better medical outcomes for the patient and her family.
Disclaimer: I have no financial interest in Univfy or their services.
© 2012, Carole. All rights reserved.