Beyond black-box models: explainable AI for embryo ploidy prediction and patient-centric consultation

Area of Science:

• Reproductive Medicine
• Artificial Intelligence in Healthcare
• Genetics and Genomics

Background:

• Predicting embryo ploidy status is crucial for successful in vitro fertilization (IVF).
• Traditional methods often lack accuracy and transparency in embryo assessment.
• Explainable Artificial Intelligence (XAI) offers a potential solution for improving predictive models.

Purpose of the Study:

• To evaluate the effectiveness of an XAI model in predicting embryo ploidy status.
• To enhance the accuracy and transparency of embryo selection in IVF.
• To identify key factors influencing embryo ploidy using AI-driven insights.

Main Methods:

• Retrospective analysis of 1908 blastocyst embryos with ploidy status, morphokinetic, and clinical data.
• Trained six machine learning (ML) models, including Random Forest (RF), on distinct embryo datasets.
• Interpreted ML model performance using XAI techniques: SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME).

Main Results:

• The RF model demonstrated superior performance, achieving an accuracy of 0.749 and AUC of 0.808 for all-grade embryos (AGE).
• External validation showed RF accuracy of 0.714 and AUC of 0.750.
• SHAP analysis identified maternal age, paternal age, time to blastocyst, and day 5 morphology grade as significant predictors.

Conclusions:

• XAI algorithms can significantly enhance the accuracy and transparency of embryo ploidy prediction.
• The developed model aids in optimizing embryo selection for patient-centric consultation.
• XAI provides reliable and transparent insights, improving decision-making in IVF.