Accurate selection of embryos with high likelihood of blastocyst formation remains a central challenge in in vitro fertilization (IVF). This study evaluated a machine-learning fusion approach that combines detailed zygote morphokinetics with routine embryo assessment to predict blastocyst outcomes. The primary objective was to perform both internal and external validation of a gradient-boosted decision tree (GBDT) fusion model and compare its discrimination with senior embryologist consensus.
This retrospective cohort included 631 normally fertilized zygotes collected from 218 treatment cycles. The model was trained and internally evaluated using 5-fold cross-validation. For external validation, the authors used a public dataset of 523 embryos with recorded blastocyst outcomes. Performance metrics reported included area under the ROC curve (AUC), area under the precision-recall curve (AUPRC), F1 score, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV).
The fusion model is a GBDT ensemble that integrated two classes of inputs: 84 zygote morphokinetic parameters and 8 conventional embryo assessment features. The source reports that the model was constructed to be interpretable and to serve as a decision-support tool for embryo selection. Specific algorithmic hyperparameters, feature engineering steps, and exact variable lists were not detailed in the preprint beyond the counts of morphokinetic and conventional features.
Using 5-fold cross-validation on the 631 zygotes, the GBDT fusion model achieved an internal AUC of 0.78 (95% CI 0.74–0.82). Additional internal performance metrics were: AUPRC 0.72, F1 score 0.73, sensitivity 0.74, specificity 0.77, PPV 0.72, and NPV 0.79. These results indicate good discrimination and balanced predictive performance across common classification metrics within the development dataset.
When applied to the external public dataset of 523 embryos, the model demonstrated acceptable generalizability with an AUC of 0.76 (95% CI 0.71–0.81). The preprint presents this as evidence that the fusion approach maintains performance on an independent dataset, supporting potential broader applicability beyond the development center.
Discrimination by the model was directly compared with senior embryologist consensus. The model significantly outperformed embryologist consensus: embryologist AUC was reported as 0.70, and the difference in AUCs favored the model with P < 0.001 by the DeLong test. Agreement between the model’s predictions and embryologist consensus was quantified with Cohen’s kappa and reported as 0.56, interpreted as moderate agreement.
Decision curve analysis was used to evaluate clinical net benefit across a range of decision thresholds. The authors report that the model provided net benefit at threshold probabilities between 0.15 and 0.55, suggesting that using the model to guide selection decisions could improve clinical decision-making relative to treating all or none across these probability thresholds.
Ethics approval was obtained from the Institutional Review Board of Guangdong Provincial People’s Hospital (approval number GDREC2022-045). The authors declare that all necessary patient/participant consent and reporting guidelines were followed. Data generated in the study are available from the authors upon reasonable request. Funding was provided by the Guangdong Medical Science and Technology Research Fund (Grant No. A2023001). The authors declared no competing interests.
A GBDT fusion model that integrates zygote morphokinetics with conventional embryo assessment features demonstrated good discrimination for blastocyst prediction, with internal AUC 0.78 and external AUC 0.76. The model outperformed embryologist consensus in discrimination and showed moderate agreement with clinicians. Decision curve analysis suggested clinical net benefit across a mid-range of threshold probabilities. The authors propose the model as an interpretable decision-support tool to aid embryo selection in IVF practice.
The preprint provides summary performance metrics and validation results but does not detail some methodological elements in the presented abstract: for example, exact lists of the 84 morphokinetic parameters, model hyperparameters, training procedures, or calibration metrics were not reported in the source abstract. Further methodological details and prospective evaluation would be needed to assess implementation logistics and real-world impact.