Construction of a Standardized Time-Lapse Imaging Database and a Gradient Boosting Ensemble Framework for Integrating Zygote Morphokinetic Parameters with Conventional Embryo Assessment

In vitro fertilization (IVF) laboratories equipped with time-lapse incubators generate vast quantities of sequential embryo images, yet the absence of standardized, annotated databases impedes the development of reproducible computational tools for embryo assessment. Here we describe the standardized time-lapse imaging database built upon prior research ground-work comprising 631 two-pronuclear (2PN) zygotes from 218 treatment cycles performed at Guangdong Provincial People’s Hospital (2020–2023), together with a gradient boosting decision tree (GBDT) ensemble framework designed to fuse heterogeneous data types for blastocyst outcome prediction. Each embryo record integrates 84 zygote-stage morphokinetic parameters extracted from EmbryoScope time-lapse sequences via a previously validated convolutional neural network segmentation pipeline with 8 conventional embryo assessment features recorded at cleavage and blastocyst stages according to the Istanbul consensus. The fusion framework employs LightGBM with equal-weight initialization and iterative residual-decreasing training, augmented by recursive feature elimination and nested five-fold cross-validation. Ablation experiments demonstrate that the full model (AUC = 0.78) outperforms morphokinetics-only (AUC = 0.71) and conventional-only (AUC = 0.65) configurations, confirming that zygote-stage temporal dynamics carry complementary information beyond standard morphological grading. SHAP analysis identifies cytoplasmic area slope, zona pellucida grayscale trend, and pronuclear fading time as the three most influential predictors. The database and fusion methodology provide a reproducible framework for integrating time-series imaging features with categorical clinical assessments in reproductive medicine.