Explainable Artificial Intelligence in Assisted Reproductive Technology: Bridging Prediction and Clinical Judgment

Background/Objectives: Artificial intelligence (AI) models are increasingly applied across the assisted reproductive technology (ART) workflow, including male-factor assessment, ovarian stimulation, endometrial receptivity evaluation, embryo selection and prediction of pregnancy outcomes. However, many systems remain difficult to interpret, raising concerns regarding transparency, clinical integration and patient communication. Explainable artificial intelligence (XAI) aims to address these limitations by making model behavior more accessible to clinicians and embryologists. This review aimed to provide a narrative, concept-driven synthesis of how XAI has been implemented in ART, to critically examine methodological quality and clinical relevance and to outline priorities for responsible translation into practice. Methods: A structured narrative review was conducted using PubMed/MEDLINE as the primary database, supplemented by targeted reference-list screening of key primary studies and recent cross-disciplinary reviews relevant to AI in ART. Studies were curated and classified according to stage of the ART workflow, data modality, model family, explanation technique and validation strategy. Methodological features, performance reporting and implementation considerations were qualitatively appraised. Results: Most XAI applications in ART fall into two dominant categories: (i) feature-attribution methods such as SHAP and LIME applied to tabular clinical and laboratory data and (ii) saliency-based approaches, including Grad-CAM and related techniques, applied to embryo and ultrasound imaging. These methods can improve transparency and support counselling by clarifying which variables or image regions influence predictions. However, the majority of studies are retrospective and single centre, with limited external validation and heterogeneous outcome definitions, often prioritising clinical pregnancy over live birth. Calibration, decision-analytic evaluation and prospective assessment remain uncommon. XAI outputs are frequently interpreted as biologically causal despite being derived from observational data, highlighting the need for cautious clinical framing. Conclusions: XAI in ART has progressed from proof-of-concept demonstrations to early clinically oriented tools, but robust validation, standardised reporting and thoughtful workflow integration are still needed. Explanations can enhance auditability and communication, yet they do not compensate for methodological weakness. Future progress will depend on higher-quality multi-centre data, evaluation beyond discrimination metrics and governance frameworks that ensure transparency, fairness and sustained performance in real-world practice.