Machine Learning-Based Enhanced Embryo Selection Using Temporal Feature Extraction and Rule-Based Labeling Model With Decision Trees

Automated embryo quality prediction remains a critical challenge in Assisted Reproductive Technology (ART) due to the subjective nature of manual grading and the lack of methods that integrate temporal developmental patterns. Embryo imaging datasets are scarce and typically limited in size. Existing approaches often neglect the time-series behavior of embryos, discard unlabeled embryos, and overlook the predictive utility of unlabeled data. Deep learning models require massive amounts of data to learn features. To overcome these limitations, we introduce an integrated framework that combines NASNet-based stage annotation, forward-filled time-aligned features, decision tree-based label imputation, and a fine-tuned SMOTE-augmented XGBoost classifier. To address the issue of missing labels, we introduce a novel rule-based decision-tree imputation strategy. It leverages domain-driven criteria and temporal features to infer missing labels, thereby enhancing dataset completeness for deep learning models. This integrated approach captures temporal developmental patterns and morphological quality while addressing class imbalance and label sparsity. Our model achieves superior predictive performance, with an accuracy of 90.1%, a macro F1-score of 0.844, and a weighted F1-score of 0.902, outperforming the baseline XGBoost and standard SMOTE variants. Specifically, without feature engineering or appropriate handling of missing values, the macro F1-score was only 0.479; with feature engineering alone without imputation, it is reduced to 0.356; however, with feature engineering combined with decision tree imputation, performance improved to a macro F1 of 0.844. Additionally, SHAP-based interpretability analysis reveals biologically meaningful features contributing to predictions. These results demonstrate the potential of our framework to support objective, interpretable, and data-efficient embryo selection in ART workflows.