AI-EMPOWERED EMBRYO ANALYSIS CAN ACHIEVE CONSISTENT, HIGH PREDICTIVE PERFORMANCE ACROSS LABORATORIES WITH HETEROGENEOUS GRADING STANDARDS AND CUTOFFS, SUPPORTING GRADING ALIGNMENT AMONG EMBRYOLOGISTS.

OBJECTIVE:

To evaluate whether an AI-based embryo scoring system provides consistent prediction of fetal heart tone (FHT) across IVF clinics, compared with morphology-based grading that may vary by laboratory.

MATERIALS AND METHODS:

This retrospective multicenter cohort study included 391 blastocyst transfer cycles with known FHT outcomes from three IVF clinics between January 2022 and August 2025. Clinics A (n=37) and B (n=242) provided both morphology grades and outcomes, while Clinic C (n=112) provided outcomes only.

The AI model predicted FHT probability from a single static Day-5 blastocyst image. Morphology grading was converted into a 1–36 (3CC to 6AA) ordinal score benchmarked from a literature (Table). Predictive performance was assessed using AUROC and accuracy across multiple cutoffs, with both clinic-specific and pooled analyses performed.

RESULTS:

Using a fixed cutoff of 5, the AI-based scoring system demonstrated consistent performance across clinics. In Clinic A, AUROC was 0.644 with accuracy of 0.676. In Clinic B, AUROC was 0.694 with accuracy of 0.657. In Clinic C, AUROC was 0.763 with accuracy of 0.696.

In contrast, morphology-based grading showed lower performance and strong dependence on clinic-specific thresholds. When Clinics A and B were pooled and evaluated at 23 (5BB), AUROC was 0.638 with accuracy of 0.570. Optimal cutoffs differed by clinic: Clinic A achieved the highest AUROC (0.704) at 27 (5BA) but with low accuracy (0.484), while Clinic B showed optimal performance at 22 (4BB) (AUROC 0.656, accuracy 0.607).

The observed variability in optimal morphology-based thresholds across clinics may reflect both inter-observer variability in grading and differences in laboratory conditions influencing embryo morphology. In contrast, AI-based scoring maintained stable predictive performance using a single cutoff across heterogeneous clinical settings.

CONCLUSIONS:

AI-based embryo scoring provides consistent and generalizable prediction of FHT across IVF clinics. Unlike morphology-based grading, which is influenced by subjective interpretation and laboratory-specific conditions, AI enables a more objective and standardized approach to embryo evaluation.

IMPACT STATEMENT:

AI-based embryo scoring may support harmonization of embryo selection across IVF laboratories by reducing variability associated with subjective grading and center-specific practices.