AI will revolutionize assisted reproductive technology (if we work together)

Consider This
AI will revolutionize assisted reproductive technology (if we work together)


Sahil Gupta, M.B.B.S.

Aveya Fertility New Delhi, Delhi INDIA

Consider This:

Infertility rates in the United States range from 6 to 18%; however, couples receive successful infertility treatment at rates of less than 1%.1,2 The major barriers to the broader use of Assisted Reproductive Technology (ART) are lack of access, high cost, complexity of treatment, and low success rates.3,4 Despite declining fertility rates in Western countries5, relatively few resources have been dedicated to reproductive research. For example, NIH funds earmarked for contraception outpaces infertility research 4 to 5-fold. Indeed, reproductive research funding trails many other diseases of smaller prevalence. The field is ripe for innovation.

Andrew Moore, a former Dean of the School of Computer Science at Carnegie Mellon University defined Artificial Intelligence (AI) as “the science and engineering of making computers behave in ways that, until recently, we thought required human intelligence.” A branch of AI, Machine Learning (ML), are systems that can learn from vast amounts of clinical, demographic, pathology, laboratory and imaging data to make connections and recommendations that humans cannot easily detect.6 In essence, ML is a set of algorithms that learn how to learn by seeing patterns and associations in large datasets. The learning can be guided or autonomous, but it produces outputs that can be used to guide decisions—including health care decisions.

AI and ML are rapidly changing the practice of medicine across various disciplines. Major inroads have already been made in disciplines where pattern recognition and classification are integral to the practice such as dermatology7, radiology8, and pathology.9 As a field, reproduction science has been slow to pursue opportunities in AI for reasons that are not at all clear. AI has remarkable potential to overcome the barriers of cost, access, and low success rates.

Consider the highly manual and labor-intensive processes of ART as it is today. Success rates depend on several variables. Some variables are patient-specific and (likely) uncontrollable, but many others are rooted in the process, including sperm, oocyte, and embryo selection to fertilization to implantation. Lack of automation leads to high inter-user variability. Indeed, gifted embryologists can be quite successful after years of training and practice; however, the learning curve and inconsistency across providers are rate-limiting. These factors are also a source of considerable costs for the practice. Automating and streamlining the entire process should reduce overhead costs to fertility practices and increase access and reduce costs for patients. Innovation does not reduce revenue for clinicians—quite the opposite—increased access, more efficient processing, and better outcomes should increase patient volume and revenue (and vastly improve patient outcomes) while reducing manual workload.

It is useful to highlight some examples of how AI can be applied to the practice of ART. For instance, researchers have had nascent successes in using AI to identify and characterize the most viable oocytes and embryos. Normally, embryologists select oocytes and embryos based on a subjective set of criteria, often developed from personal experience rather than evidence-based sources. To standardize, formalize and improve the selection process, researchers created and tested an AI system on two data sets of 269 oocytes and 269 corresponding embryos from 104 women.10 The AI was able to successfully identify and classify oocyte/embryo quality using the information it had learned through previous training.

Indeed, as the dataset expands and additional ML continues, the AI can hone its accuracy and predictive prowess. The algorithm gets “rewarded” for choosing features that are ultimately associated with better outcomes (live births, successful implantations, etc.). Put another way, the algorithm mathematically weights the features that lead to better outcomes. Conversely, the algorithm is “penalized” for identifying features associated with poorer outcomes (i.e. lower mathematical weight). In time, unsupervised, “deep learning” AI can detect patterns and features that the original programmers had not overtly considered or that embryologists may not use to subjectively assign quality. Yet since the AI is reinforced by outcomes, positive and negative, it creates a sort of inherent, evidence-based selection strategy. Moreover, the fully trained AI could be standardized, commercialized, and marketed for use in fertility clinics. This would take the guesswork out of one of the major sources of variability in the ART process.

One could also envision AI used in a similar fashion to characterize spermatocyte quality. Computer-aided sperm analysis (CASA) systems are used in research and have been adopted in some clinics. CASA assesses motile percentage and kinematic parameters at the population level. Standard parameters include amplitude of lateral head displacement, average path velocity, beat cross frequency, curvilinear velocity, straight-line velocity, straightness, and linearity.11 However, when a relatively simple AI algorithm was trained on 2817 spermatocytes from 18 individuals and tested, the AI was able to correctly classify sperm motility into five classes: progressive vigorous, intermediate vigorous, or hyperactivated vigorous, slow non-vigorous, or weakly motile non-vigorous.11 Overall accuracy was 89.9%.

Researchers have also had early success in using ML to digitally isolate sperm cell heads and characterize sperm head morphology as “good” or “bad.”12 The automatic analysis was trained on over 1,400 human sperm cells from 8 donors and reached precisions of 88% and higher.12

Large datasets (i.e. “big data”) allow ML to learn and develop best practices that can be applied to ART. For example, 80% of one large dataset can be used to train the AI while the remaining 20% is used to validate the model.13 Then a second large dataset can be used to test the AI algorithm and compare it to historical diagnostic pathways, assessments, treatments, and outcomes.13 Once trained, the AI could help the fertility specialist in several ways. AI can be used to hone ovarian stimulation protocols. The AI can crawl through thousands of electronic medical records to make associations between ovarian stimulation parameters and outcomes. In a similar way, AI could also help predict IVF outcome to the patient-level. Data mining has been used to identify factors that predict IVF outcome such as the age of woman, the number of the developed embryos and the serum estradiol level on the day of human chorionic gonadotropin administration.14 As these technologies are developed, the precision should increase to the point of being able to reliably assign an outcome probability on a patient-by-patient basis.

While the widespread use of electronic medical records will help pave the way for data mining and AI applications, the high variability of stimulation and embryology techniques across laboratories is major barrier to ML. While newer AI algorithms can partially compensate for missing data, all ML systems work best when they can learn on vast, complete, codified data. Until reproductive specialists adopt a common clinical language and standard data acquisition criteria, data mining cannot occur to the degree required for off-the-shelf ART applications. Thus, the near-term will likely be an iterative process. AI can begin to learn from partial, varied data and provide limited insights—insights limited by the quality of the data from which they learn. Reproductive specialists can begin to standardize their systems as collective knowledge grows. Comprehensive note-taking, detailed outcomes reporting, and routine collection of high quality imaging, can accelerate this innovation. As such, all fertility specialists can take part in the AI revolution in ART.



  1. Chandra A, Copen CE, Stephen EH. Infertility and impaired fecundity in the United States, 1982-2010: data from the National Survey of Family Growth. Natl Health Stat Report. 2013(67):1-18, 11 p following 19.
  2. Thoma ME, McLain AC, Louis JF, et al. Prevalence of infertility in the United States as estimated by the current duration approach and a traditional constructed approach. Fertil Steril. 2013;99(5):1324-1331 e1321. 10.1016/j.fertnstert.2012.11.037
  3. Paulson RJ, Fauser B, Vuong LTN, Doody K. Can we modify assisted reproductive technology practice to broaden reproductive care access? Fertil Steril. 2016;105(5):1138-1143. 10.1016/j.fertnstert.2016.03.013
  4. De Geyter C, Calhaz-Jorge C, Kupka MS, et al. ART in Europe, 2014: results generated from European registries by ESHRE: The European IVF-monitoring Consortium (EIM) for the European Society of Human Reproduction and Embryology (ESHRE). Hum Reprod. 2018;33(9):1586-1601. 10.1093/humrep/dey242
  5. Population GBD, Fertility C. Population and fertility by age and sex for 195 countries and territories, 1950-2017: a systematic analysis for the Global Burden of Disease Study 2017. Lancet. 2018;392(10159):1995-2051. 10.1016/S0140-6736(18)32278-5
  6. Obermeyer Z, Emanuel EJ. Predicting the Future - Big Data, Machine Learning, and Clinical Medicine. N Engl J Med. 2016;375(13):1216-1219. 10.1056/NEJMp1606181
  7. Hogarty DT, Su JC, Phan K, et al. Artificial Intelligence in Dermatology-Where We Are and the Way to the Future: A Review. Am J Clin Dermatol. 2020;21(1):41-47. 10.1007/s40257-019-00462-6
  8. Hosny A, Parmar C, Quackenbush J, Schwartz LH, Aerts H. Artificial intelligence in radiology. Nat Rev Cancer. 2018;18(8):500-510. 10.1038/s41568-018-0016-5
  9. Colling R, Pitman H, Oien K, et al. Artificial intelligence in digital pathology: a roadmap to routine use in clinical practice. J Pathol. 2019;249(2):143-150. 10.1002/path.5310
  10. Manna C, Nanni L, Lumini A, Pappalardo S. Artificial intelligence techniques for embryo and oocyte classification. Reprod Biomed Online. 2013;26(1):42-49. 10.1016/j.rbmo.2012.09.015
  11. Goodson SG, White S, Stevans AM, et al. CASAnova: a multiclass support vector machine model for the classification of human sperm motility patterns. Biol Reprod. 2017;97(5):698-708. 10.1093/biolre/iox120
  12. Mirsky SK, Barnea I, Levi M, Greenspan H, Shaked NT. Automated analysis of individual sperm cells using stain-free interferometric phase microscopy and machine learning. Cytometry A. 2017;91(9):893-900. 10.1002/cyto.a.23189
  13. Komorowski M, Celi LA, Badawi O, Gordon AC, Faisal AA. The Artificial Intelligence Clinician learns optimal treatment strategies for sepsis in intensive care. Nat Med. 2018;24(11):1716-1720. 10.1038/s41591-018-0213-5
  14. Hafiz P, Nematollahi M, Boostani R, Namavar Jahromi B. Predicting Implantation Outcome of In Vitro Fertilization and Intracytoplasmic Sperm Injection Using Data Mining Techniques. Int J Fertil Steril. 2017;11(3):184-190. 10.22074/ijfs.2017.4882