Robust and generalizable embryo selection based on artificial intelligence and time-lapse image sequences.

Area of Science:

• Embryology
• Artificial Intelligence
• Reproductive Medicine

Background:

• Embryo selection is crucial for in vitro fertilization (IVF) success.
• Artificial intelligence (AI) and deep learning offer potential for automating embryo assessment.
• Limited research exists on the generalizability and subgroup performance of AI embryo selection models.

Purpose of the Study:

• To investigate the performance of a deep learning-based embryo selection model using time-lapse imaging.
• To evaluate the model's generalizability across different IVF clinics and patient subgroups.
• To assess the correlation between AI predictions and traditional morphokinetic parameters.

Main Methods:

• A deep learning model was trained on a large dataset (115,832 embryos) from 18 IVF centers.
• The model utilized time-lapse image sequences for embryo scoring.
• Performance was evaluated using area under the curve (AUC) on independent test sets, including clinic hold-out and subgroup analyses.

Main Results:

• The AI model achieved an AUC of 0.67 for predicting implantation of known data (KID) embryos and 0.95 for all embryos.
• The model demonstrated generalizability to new clinics (AUC 0.60-0.75) and performed consistently across patient age and clinical condition subgroups (AUC 0.63-0.69).
• Model predictions correlated positively with blastocyst grading and negatively with direct cleavages, performing comparably to manual selection.

Conclusions:

• The fully automated iDAScore v1.0 model provides reliable embryo selection based on time-lapse imaging.
• The AI model generalizes well to new clinical settings and diverse patient populations.
• Automated embryo scoring reduces manual effort and eliminates observer variability, enhancing IVF efficiency and outcomes.