Study question

Does a CNN trained to predict fetal heartbeat pregnancy from day-5 fresh blastocyst images also predict biochemical pregnancy, and how does outcome definition affect performance?

Summary answer

The CNN predicted both outcomes, with higher discrimination for fetal heartbeat pregnancy and lower but significant discrimination for biochemical pregnancy, indicating outcome-dependent model performance.

What is known already

Pregnancy outcomes in assisted reproduction are variably defined, including biochemical pregnancy based on urine or serum β-human chorionic gonadotropin (b-hCG) or as clinical pregnancy confirmed by fetal heartbeat (FHT). Biochemical pregnancy reflects early implantation but does not necessarily progress to sustained embryonic development. Many embryo prediction models are developed using FHT-based endpoints, whereas biochemical pregnancy remains widely used in the clinical setting. However, within the same cohort, it remains unclear how outcome definition influences model discrimination when a single image-based deep learning model is applied, and whether biological interpretation differs between biochemical and FHT pregnancy.

Study design, size, duration

This retrospective multicentre cohort study included fresh day-5 embryo transfer cycles performed between May 2018 and July 2024, based on oocyte pickup dates, across three centres. A total of 251 embryo transfer cycles were analysed, comprising 406 transferred embryos. Single and multiple embryo transfers were included; multiple-embryo transfers were limited to cycles with known implantation data (KID), where all embryos resulted in pregnancy or non-pregnancy.

Participants/materials, setting, methods

A convolutional neural network (CNN) model was trained on day-5 fresh blastocyst images to predict the probability of fetal heartbeat (FHT). Training used 4,866 fresh embryo images collected from 17 IVF clinics between July 2011 and August 2025. Biochemical pregnancy was defined as serum b-hCG ≥100, and FHT pregnancy was confirmed by ultrasound. Predictive discrimination was assessed using AUC and Youden-derived cut-offs. Differences in AUC were evaluated using bootstrap resampling (1,000 iterations).

Main results and the role of chance

Among 406 embryos, biochemical pregnancy occurred in 20.0% and FHT pregnancy in 21.2%. The embryo viability score was significantly higher in biochemical pregnancy than in non-pregnancy cases (mean±SD: 5.46±2.19 vs 4.16±2.28; p < 0.001).

For biochemical pregnancy prediction, the score demonstrated meaningful discrimination (AUC=0.663). Each one-unit increase in score was associated with higher odds of biochemical pregnancy (odds ratio 1.30; 95% CI 1.16–1.46; p < 0.0001). At a Youden-derived cut-off >5.05, sensitivity was 0.69 and specificity was 0.60.

Discrimination was higher for FHT pregnancy (AUC=0.761). Bootstrap analysis confirmed a significant difference in AUC between FHT and biochemical pregnancy outcomes (ΔAUC=0.098; 95% CI 0.046–0.150; p = 0.00023).

Exploratory analyses showed that ICM and TE grades were significantly associated with FHT pregnancy, whereas biochemical pregnancy was consistently associated with TE grade only. This pattern may suggest that early implantation is primarily driven by TE-related features, while progression to FHT requires additional developmental competence reflected by ICM quality.

Limitations, reasons for caution

This retrospective analysis included fresh day-5 transfers from three IVF centres. Definitions of biochemical pregnancy (b-hCG ≥100) and reporting practices may vary across settings, limiting generalisability. External validation in larger, international cohorts, including frozen embryo transfer cycles, is warranted.

Wider implications of the findings

Embryo models developed to predict fetal heartbeat pregnancy may also inform biochemical pregnancy, but interpretation requires consideration of differences in the morphological features associated with each outcome. Biochemical pregnancy was primarily associated with trophectoderm features, whereas both inner cell mass and trophectoderm features contributed to fetal heartbeat prediction.