Study question Can device-independent artificial intelligence (AI) models accurately predict blastocyst formation and quality by embryonic day 3 using time-lapse images and age?

Summary answer The AI-based prediction models achieved receiver operating characteristic (ROC) area under the curve (AUC) ≥0.87 for predicting blastocyst formation, good blastocysts, and poor blastocyst + arrested embryos, demonstrating high accuracy at embryonic day 3.

What is known already Previous blastocyst prediction models from embryonic day 3 time-lapse images and parameters have achieved moderate accuracy (ROC AUCs: 0.73–0.82). These models require labor-intensive manual parameter annotation or exhibit device dependence, limiting their generalizability and clinical usefulness.

Study design, size, duration This was a retrospective, multicenter study using 7,111 eligible human embryos cultured in four types of time-lapse incubators across four facilities from May 2018 to December 2022. Additionally, 23,852 time-lapse and stereomicroscopic images of embryos during the same period were collected and annotated using 17-class supervised classification labels for the AI auto-annotator.

Participants/materials, setting, methods Embryo culture time-lapse images and clinical data were retrospectively collected from Kyoto University Hospital and its cooperating facilities. This study included 2PN embryos cultured individually per well up to embryonic day 5 or 6. Degenerated embryos and those with poor Veeck grades (4 and 5) were excluded. Time-lapse images were recorded every 15 minutes between 24- and 64-hours post-insemination and preprocessed. Embryos with over 5% image loss or preprocessing errors were excluded. The final dataset was used to construct AI models for blastocyst formation and quality prediction. The dataset was split into 80% for training and cross-validation and 20% for independent testing. A pre-trained image classification AI (NASNet-A Large) was fine-tuned as AI auto-annotator, using 17-class supervised classification dataset including cell stages and Veeck grades. The AI auto-annotator converted time-lapse images into time-series labels. These labels, the proportions of each stage and grade, and the age at egg retrieval were analyzed using a gradient boosting algorithm (XGBoost) to predict blastocyst formation and quality. The performance of models was evaluated using ROC and precision-recall analyses with 5-fold validation and independent testing.

Main results and the role of chance The AI auto-annotator achieved 85% accuracy in classifying embryo images into 17 classes, with 95% cell stage classification accuracy. The blastocyst formation prediction model had an ROC AUC of 0.86 for mean cross-validation and 0.87 for test. The good blastocyst formation prediction model had a ROC AUC of 0.88 for test. The poor blastocyst + arrested embryo prediction model had a ROC AUC of 0.87 for test and achieved an AUC of 0.90 for precision recall, meeting the target accuracy for clinical decision-making. The performance of models was evaluated through five-fold cross-validation and independent testing. Subgroup analyses revealed minimal differences between facilities and age groups, demonstrating the generalizability of the model.

Limitations, reasons for caution The predictive models used Gardner criteria, widely adopted but prone to inter-rater variability, potentially affecting consistency. Blastocyst evaluations at each facility were unmodified, and single-rater evaluations were avoided to prevent domain shift. The exclusion of embryos with defects or poor quality (Veeck grades 4 and 5) may limit the model’s generalizability. Additionally, as this was a retrospective study, prospective validation is required to confirm the findings.

Wider implications of the findings This versatile AI model enables more accurate embryo-by-embryo decisions by embryonic day 3, potentially guiding selection between continuing blastocyst culture versus early transfer/freezing, and thus improving individualized patient care in ART.

Study funding/competing interest(s) This work was supported by JSPS KAKENHI grant number JP23K08821. The authors have no conflicts of interest to declare.

Trial registration number N/A.

Capsule Artificial intelligence models accurately predicted blastocyst formation and quality using time-lapse images and age on embryonic day 3, supporting clinical decision-making regarding blastocyst culture or early embryo transfer.

The authors have declared no competing interest.

This work was supported by JSPS KAKENHI grant number JP23K08821.

I confirm all relevant ethical guidelines have been followed, and any necessary IRB and/or ethics committee approvals have been obtained.

Yes

The details of the IRB/oversight body that provided approval or exemption for the research described are given below:

This study was conducted in accordance with the principles embodied in the Declaration of Helsinki and was approved by the Ethics Committee of Kyoto University Graduate School and Faculty of Medicine (approval number R3320). Informed consent was obtained in the form of opt-out through the websites of facilities.

I confirm that all necessary patient/participant consent has been obtained and the appropriate institutional forms have been archived, and that any patient/participant/sample identifiers included were not known to anyone (e.g., hospital staff, patients or participants themselves) outside the research group so cannot be used to identify individuals.

Yes

I understand that all clinical trials and any other prospective interventional studies must be registered with an ICMJE-approved registry, such as ClinicalTrials.gov. I confirm that any such study reported in the manuscript has been registered and the trial registration ID is provided (note: if posting a prospective study registered retrospectively, please provide a statement in the trial ID field explaining why the study was not registered in advance).

Yes

I have followed all appropriate research reporting guidelines, such as any relevant EQUATOR Network research reporting checklist(s) and other pertinent material, if applicable.

Yes

The AI code and required weights for this study are available at GitHub (https://github.com/KU-ObGy-ART/blastocyst-prediction/). Owing to the nature of this research, the participants of this study did not agree that their data, including time-lapse images, should be shared publicly. Thus, images and clinical features of the fine-tuning and blastocyst prediction datasets are not available.