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

Technical Summary: Construction of a Standardized Time-Lapse Imaging Database and a Gradient Boosting Ensemble Framework

Problem Statement
While time-lapse incubators in IVF laboratories generate vast quantities of sequential embryo images, the field lacks standardized, annotated databases that link image sequences with clinical metadata and outcome labels. This absence hinders the development of reproducible computational tools. Furthermore, Assisted Reproductive Technology (ART) data is inherently heterogeneous, comprising continuous morphokinetic time series, ordinal morphological grades, and categorical patient variables. Existing public resources often lack comprehensive outcome labels or zygote-stage kinetic features, and standard deep learning pipelines require modality-specific architectures and large training sets that may not be available. Consequently, there is a need for a principled approach to fuse these disparate data types to improve blastocyst outcome prediction beyond static morphological grading.

Methodology
The study presents a two-pronged approach: the construction of a standardized database and the development of a specific machine learning framework.

• Database Construction: The authors compiled a dataset of 631 two-pronuclear (2PN) zygotes from 218 treatment cycles (2020–2023) at Guangdong Provincial People's Hospital. The database integrates:

  • Clinical Data: 12 patient-level variables (e.g., maternal age, BMI, stimulation protocol).
  • Conventional Assessment: 8 embryo-level features recorded at Day 2, Day 3, and Day 5/6 according to the Istanbul consensus (cell count, fragmentation, symmetry, blastocyst grades).
  • Morphokinetic Parameters: 84 features extracted from EmbryoScope time-lapse sequences using a previously validated CNN segmentation pipeline (US Patent US11210494B2). These features cover temporal dynamics of the zona pellucida, cytoplasm, and pronuclei (e.g., area slopes, grayscale trends, fading times) over the first 18 hours post-insemination.
  • Outcome: A binary label indicating whether a usable blastocyst was formed by Day 5 or Day 6.

• Gradient Boosting Decision Tree (GBDT) Ensemble Framework: To handle the heterogeneous data, the authors employed LightGBM with a specific design principle: equal-weight initialization.

  • Initialization: At the first boosting iteration, morphokinetic (continuous) and conventional (ordinal/categorical) features are assigned comparable base-learner influence to prevent high-variance continuous features from dominating initial tree splits.
  • Training: The model utilizes iterative residual-decreasing training. Subsequent boosting rounds adaptively allocate model capacity based on the information content of each feature modality, guided by residual gradients.
  • Feature Selection: Recursive Feature Elimination (RFE) was applied within a nested five-fold cross-validation scheme. Starting with 92 features (84 morphokinetic + 8 conventional), the process retained 23 features (19 morphokinetic, 4 conventional) to optimize performance.
  • Evaluation: Model performance was assessed using the Area Under the Curve (AUC) and Brier score, with statistical comparisons made via the DeLong test. Feature importance was quantified using SHAP (SHapley Additive exPlanations) values.

Key Results

• Model Performance: The full fusion model (integrating both data types) achieved an AUC of 0.78 (95% CI: 0.74–0.82). This significantly outperformed the morphokinetics-only model (AUC = 0.71, P = 0.003) and the conventional-only model (AUC = 0.65, P < 0.001).
• Complementary Information: The results confirm that zygote-stage temporal dynamics provide predictive information complementary to standard morphological grading, as evidenced by the significant performance gap between the conventional-only and full fusion models.
• Feature Importance: SHAP analysis identified the three most influential predictors as:
  1. Cytoplasmic area slope (rate of cytoplasmic volume change).
  2. Zona pellucida grayscale intensity trend.
  3. Pronuclear fading time.
     The top-ranked conventional feature was Day 3 cell count.
• Data Quality: 92.9% of the 84 morphokinetic parameters were successfully computed for all embryos. Missing values were imputed using stratum-specific medians.

Significance and Claims
The authors position this work primarily as a data resource and methodological blueprint. They explicitly state that this is not a clinical validation study and does not compare model performance against embryologist assessments. The significance of the paper lies in:

1. Providing a standardized, reproducible framework for integrating time-series imaging features with categorical clinical assessments in reproductive medicine.
2. Demonstrating that an equal-weight initialization strategy in GBDT ensembles effectively fuses heterogeneous ART data types.
3. Establishing that zygote-stage morphokinetics carry complementary information that enhances blastocyst formation prediction beyond static morphological grades.

The database and methodology are intended to facilitate reproducible research on zygote morphokinetics and enable the development of more robust embryo selection tools, subject to future clinical validation.