| 摘要: | The adoption of artificial intelligence (AI) in predicting embryo ploidy status represents an emerging trend, offering a non-invasive approach to selecting the most viable embryo for transfer. Nevertheless, current models exhibit suboptimal performance, and their decision-making processes are challenging to decipher. As a response, we propose an integrated model that combines time-lapse and clinical data to enhance predictive accuracy. Additionally, we aim to incorporate new explainable artificial intelligence (XAI) techniques to instill confidence in human users and provide transparent insights into the model's decision-making process.
Our dataset comprised 1,908 embryos from the Taipei Fertility Center in Taipei, Taiwan (2020-2022). The input variables for our models encompass morphokinetic parameters, morphology grade, and 11 additional clinical variables. Six machine learning (ML) models including Random Forest (RF), Linear Discriminant Analysis (LDA), Logistic Regression (LR), Support Vector Machine (SVM), AdaBoost (ADA), and Light Gradient-Boosting Machine (LGBM) were trained using a mix of embryonic and clinical variables to predict ploidy status probabilities across three distinct datasets: high-grade embryos (HGE), low-grade embryos (LGE), and all-grade embryos (AGE). The model's performance was interpreted using XAI, including SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) techniques.
The training and model construction incorporated 1,471 embryos, all with PGT-A results. The mean maternal age was 38.5±3.85 years, and the AMH levels were recorded at 2.57±1.87 ng/mL. The RF model exhibited superior performance compared to the other five ML models, achieving an accuracy of 74.9% and an area under the curve (AUC) of 0.808 for AGE. Similarly, the RF model demonstrated an accuracy of 71.2% for LGE compared to 69.8% for HGE, indicating enhanced performance in predicting ploidy status when combining embryos with high and low grades. In the external test set (437 embryos), the RF model achieved an accuracy of 0.714 and an AUC of 0.750 (95% CI: 0.702-0.796). SHAP's feature impact analysis highlighted that maternal age, paternal age, tB, and day 5 morphology grade significantly impacted the predictive model. In addition, LIME offered specific case-ploidy prediction probabilities, revealing the model's assigned values for each variable within a finite range.
Our study underscores the significant potential of harnessing AI algorithms to aid clinicians in embryo ploidy evaluation. By integrating XAI techniques, we can gain deeper insights into the model's decision-making process. This includes discerning the ranking of variable importance and understanding the intricate interplay between these variables. Ultimately, this advancement holds promise in enhancing outcomes within IVF, offering a personalized approach on a case-by-case basis.
The adoption of artificial intelligence (AI) in predicting embryo ploidy status represents an emerging trend, offering a non-invasive approach to selecting the most viable embryo for transfer. Nevertheless, current models exhibit suboptimal performance, and their decision-making processes are challenging to decipher. As a response, we propose an integrated model that combines time-lapse and clinical data to enhance predictive accuracy. Additionally, we aim to incorporate new explainable artificial intelligence (XAI) techniques to instill confidence in human users and provide transparent insights into the model's decision-making process.
Our dataset comprised 1,908 embryos from the Taipei Fertility Center in Taipei, Taiwan (2020-2022). The input variables for our models encompass morphokinetic parameters, morphology grade, and 11 additional clinical variables. Six machine learning (ML) models including Random Forest (RF), Linear Discriminant Analysis (LDA), Logistic Regression (LR), Support Vector Machine (SVM), AdaBoost (ADA), and Light Gradient-Boosting Machine (LGBM) were trained using a mix of embryonic and clinical variables to predict ploidy status probabilities across three distinct datasets: high-grade embryos (HGE), low-grade embryos (LGE), and all-grade embryos (AGE). The model's performance was interpreted using XAI, including SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) techniques.
The training and model construction incorporated 1,471 embryos, all with PGT-A results. The mean maternal age was 38.5±3.85 years, and the AMH levels were recorded at 2.57±1.87 ng/mL. The RF model exhibited superior performance compared to the other five ML models, achieving an accuracy of 74.9% and an area under the curve (AUC) of 0.808 for AGE. Similarly, the RF model demonstrated an accuracy of 71.2% for LGE compared to 69.8% for HGE, indicating enhanced performance in predicting ploidy status when combining embryos with high and low grades. In the external test set (437 embryos), the RF model achieved an accuracy of 0.714 and an AUC of 0.750 (95% CI: 0.702-0.796). SHAP's feature impact analysis highlighted that maternal age, paternal age, tB, and day 5 morphology grade significantly impacted the predictive model. In addition, LIME offered specific case-ploidy prediction probabilities, revealing the model's assigned values for each variable within a finite range.
Our study underscores the significant potential of harnessing AI algorithms to aid clinicians in embryo ploidy evaluation. By integrating XAI techniques, we can gain deeper insights into the model's decision-making process. This includes discerning the ranking of variable importance and understanding the intricate interplay between these variables. Ultimately, this advancement holds promise in enhancing outcomes within IVF, offering a personalized approach on a case-by-case basis. |
