黎阮國慶BUI XUAN LAM2026-09-092026-09-092026-07-20https://handle.ncl.edu.tw/11296/y2kvmzhttps://203.71.86.71/handle/123456789/74537學位別：碩士 語文別：英文 指導教授：黎阮國慶 口試委員：蔡承育; 黎阮國慶; 高淑慧 授權範圍：網際網路,開放日期為2026-07-26 電子論文連結：https://handle.ncl.edu.tw/11296/y2kvmzThe decision of ovulation trigger timing represents a critical step in in vitro fertilization (IVF) and intracytoplasmic sperm injection (ICSI) treatment, as it influences the number and maturity of oocytes retrieved and may affect downstream embryological outcomes. In current clinical practice, trigger decisions are commonly guided by the diameter of one or a few leading follicles or by simplified follicle size thresholds. Nevertheless, these approaches may not fully capture the biological information contained in the entire follicular cohort on the day of trigger (DoT). As a response, we propose a distribution based and explainable artificial intelligence (XAI) framework to evaluate whether follicle size distribution on the DoT can improve the prediction and interpretation of key laboratory outcomes in Gonadotropin-releasing hormone (GnRH) antagonist ICSI cycles. Our dataset comprised 907 eligible GnRH antagonist ICSI cycles from Phuong Chau International Hospital, Can Tho, Vietnam, between 2011 and 2025. To reduce heterogeneity, the study included one eligible cycle per woman and excluded cycles with missing trigger day follicle distribution, polycystic ovary syndrome (PCOS), inconsistent records, or oocyte donation. The development cohort included 852 cycles treated between 2011 and 2024, while an independent temporal test cohort of 55 cycles from 2025 was reserved for final model evaluation. Trigger day follicle size distribution was represented using ten predefined follicle size bins ranging from 6-8 mm to 25-26 mm. The primary outcomes were total oocytes retrieved, metaphase II (MII) oocytes, and two-pronuclear (2PN) zygotes. Several statistical and machine learning (ML) models were evaluated, including negative binomial regression (NB), linear regression (LR), k-nearest neighbours regression (KNN), support vector regression (SVR), random forest (RF), gradient boosting machine (GBM), Extreme gradient boosting (XGBoost), and multilayer perceptron regression (MLP). Model performance was assessed using cross-validation (CV) within the development cohort and temporal validation in the 2025 test cohort. Model behaviour was further interpreted using permutation importance (PI), SHapley Additive exPlanations (SHAP), subgroup analyses, biological association analyses, exploratory clinical pregnancy prediction, counterfactual examples, and decision curve analysis. The linear SVR model demonstrated the most favourable balance between predictive performance, stability, and interpretability. In the temporal test cohort, the final model achieved an R² of 0.771 and a mean absolute error (MAE) of 2.518 for total oocytes retrieved, an R² of 0.684 and an MAE of 2.422 for MII oocytes, and an R² of 0.558 and an MAE of 2.453 for 2PN zygotes. Across exploratory signal mapping, PI, and SHAP analyses, follicles within the 11–18 mm range consistently emerged as the most informative window for predicting laboratory outcomes. The 11–12 mm bin showed the strongest importance for total oocytes and MII oocytes, whereas the 17–18 mm bin contributed most strongly to 2PN prediction. In contrast, follicles larger than 18 mm showed weaker and less stable predictive contributions. Biological association analyses further supported this finding, showing that a higher proportion of 11–18 mm follicles was associated with higher MII oocyte yield, while a greater burden of follicles larger than 18 mm was linked to higher trigger day progesterone concentration. In the exploratory decision curve analysis for identifying cycles at risk of low MII yield, the model incorporating the 11–18 mm follicle distribution provided greater net benefit than the clinical model alone across clinically relevant threshold probabilities, whereas the model based on follicles larger than 18 mm did not show consistent added value. Our study underscores the potential of using the full trigger day follicle size distribution, rather than relying only on leading follicle measurements, to support individualized ovulation trigger assessment in GnRH antagonist ICSI cycles. By integrating ML with XAI, this work provides both predictive estimates and clinically interpretable information about which follicle size ranges drive model behaviour. The findings suggest that the 11–18 mm follicle window contains meaningful information for predicting oocyte retrieval, oocyte maturity, and fertilization-related outcomes. However, these results should be interpreted as predictive and explanatory rather than causal. Further external validation, prospective evaluation, and assessment of downstream reproductive outcomes are needed before distribution-informed trigger day prediction can be implemented in routine clinical practice.系所名稱：國際醫學研究碩士學位學程thesis