Predicting time to live birth with deep learning embryo ranking: a novel multiple imputation approach

Area of Science:

• Reproductive Medicine
• Artificial Intelligence in Healthcare
• Embryology

Background:

• Estimating the clinical utility of embryo selection algorithms is complex due to unknown outcomes for non-transferred embryos.
• Previous studies used biased samples or constructed data, limiting reliability.

Purpose of the Study:

• To assess the clinical utility of embryo selection algorithms in predicting time to live birth (TTLB).
• To compare the predictive accuracy of algorithms against manual embryo grading.

Main Methods:

• Retrospective cohort study of 3,783 IVF treatments (2015-2022).
• Utilized multiple imputation by chained equations (MICE) to estimate outcomes for non-transferred embryos.
• Compared a deep learning algorithm's performance against manual ranking.

Main Results:

• The deep learning algorithm predicted an average TTLB of 1.68 transfers, 6.1% shorter than manual ranking (1.78 transfers).
• Algorithm performance showed an area under the receiver operating characteristic curve (AUC) of 0.633 at the population level and 0.672 at the treatment level.

Conclusions:

• Embryo selection algorithms demonstrate potential to enhance IVF treatment efficacy by improving TTLB prediction.
• These algorithms may increase live birth rates and overall success in assisted reproductive technology.
• Estimated TTLB is an approximation, influenced by factors beyond embryo quality and specific to the dataset used.