Return to Search

Artificial Intelligence (AI) technology can predict human embryo viability across multiple laboratories with varying demographics with high accuracy and reproducibility.

Authors: 1. Corresponding/Presenting Author: Matthew D. VerMilyea (Ovation Fertility, Austin/San Antonio, TX, USA), 2. Andrew Miller (Ovation Fertility, Austin, TX, USA), 3. Michelle Lane (Repromed, Adelaide, Australia; Monash IVF Group, Richmond, Australia), 4. Glen Adaniya (Ovation Fertility, Indianapolis, ID, USA), 5. Bradford Bopp (Midwest Fertility Specialists, Carmel, ID, USA), 6. Dean E Morbeck (Fertility Associates, Auckland, New Zealand), 7. Erica Behnke (Ovation Fertility, Cincinnati, OH, USA), 8. Leah Click (Institute of Reproductive Health, Cincinnati, USA), 9. Rebecca Matthews (Oregon Reproductive Medicine, Portland, USA), 10. Adelle Lim (Alpha Fertility Centre, Petaling Jaya, Malaysia), 11. Jonathan Hall (Life Whisperer, Adelaide, Australia), 12. Michelle Perugini (Life Whisperer, Adelaide, Australia), 13. Don Perugini (Life Whisperer, Adelaide, Australia), 14. Andrew Murphy (Life Whisperer, Adelaide, Australia)

Title: Artificial intelligence (AI) technology can predict human embryo viability across multiple laboratories with varying demographics with high accuracy and reproducibility.

Study question: Can artificial intelligence (AI) and computer vision provide improvement to embryo viability prediction using static 2D images of Day 5 embryos from multiple laboratories?

Summary Answer: The development of a general AI model produced 32% improvement in accuracy regarding embryo viability assessment compared to traditional embryologist morphology assessment.

What is known already: Recent studies have shown that artificial intelligence (deep learning) and computer vision can increase the efficacy of embryo selection and prediction of clinical pregnancy using images of human embryos. This automated, non-invasive approach to embryo selection can be used as a cohort ranking tool whereby embryos with certain morphological features are ranked in order of their likelihood to result in a positive pregnancy with a fetal heartbeat. A validation study of the current AI model using ~5000 images of Day 5 blastocysts resulted in an accuracy of 67.7% in identifying embryo viability by positive fetal heartbeat across two blind datasets.

Study design, size, duration: Approximately 20,000 static 2D images of Day 5 blastocysts with related pregnancy, preimplantation genetic testing for aneuploidy (PGT-A) outcomes, demographic and clinic geographical location information have been collected. Images were divided into three groups: training, validation and blind test sets. An AI model was trained, validated and tested on 2217 embryo images from Day 5 blastocysts followed by a further blind set of 286 images from a separate clinic and demographic.

Participants/materials, setting, methods: 7,847 separate traditional phase-contrast microscope images from patients undergoing fertility treatment at 12 IVF laboratories in four countries were used to train and develop the Life Whisperer’s embryo viability assessment model (LW General Model v1). Images of Day 5 blastocysts which had a blastocoel cavity and were subsequently transferred individually were used for this study. This study was determined exempt from IRB review by Sterling IRB, USA (#6467).

Main results and the role of chance: 5282 images from Monash IVF Group (Repromed, Adelaide, SA, Australia) were split into training (~74%), validation (~7%) and two blind validation (~19%) datasets. A blind validation data set is used to conduct an unbiased accuracy assessment to ensure the model is generalizable to all embryo images. The accuracy in identifying viable embryos is calculated as a percentage of the number of viable embryos (i.e. blastocysts that resulted in a successful pregnancy) identified by the AI model divided by the total number of viable embryos in the dataset. The same process applies to non-viable embryos. The total mean accuracy of the embryo viability assessment model when applied to both blind validation datasets was 74.1% for identifying viable embryos, 65.3% non-viable embryos and 67.7% total accuracy across both viable and non-viable embryos. When comparing the accuracy in identifying viable/nonviable embryos for the model versus world-leading embryologists, the AI model correctly identified embryo viability 66.7% compared to 51.0% by embryologists. After subsequent training on a mixed demographic dataset, a blind test of 74 images from three independent laboratories (Ovation Fertility Austin, San Antonio and Indianapolis), resulted in the model correctly identifying viable embryos at 84.6%, non-viable at 57.1% and total accuracy at 71.6%.

Limitations, reasons for caution: While the analysis of the blind validation datasets is small, our results provide evidence that AI can reduce the variability of morphological embryo selection across multiple laboratories and embryologists. Additional training on larger datasets is currently ongoing to further improve the generalizability of the AI model.

Wider implications of the findings: Embryo selection is heavily based on morphological/morphokentetic assessment and other invasive technologies including embryo biopsy for embryo ploidy classification. For the first time, this data represents how an AI model can be applied across a multi-centered, mixed demographic dataset resulting in ~70% overall accuracy for viable and non-viable embryo identification.

Study funding/competing interest(s): N/A