Ultra-high-performance concrete (UHPC) is increasingly utilized in bridge deck slabs due to its superior mechanical properties and durability. However, its high autogenous shrinkage and the resulting risk of early-age cracking, especially under restrained curing conditions, present significant challenges for long-term structural integrity. Recent advances in machine learning (ML) have enabled more accurate prediction and understanding of shrinkage and cracking behaviors in UHPC, facilitating optimized mix designs and mitigation strategies. This review synthesizes over 100 recent studies on ML-based prediction of shrinkage cracking in UHPC bridge decks, focusing on quantitative model performance, influential material parameters, experimental validation, and practical engineering implications. Ensemble models such as XGBoost, Random Forest, and hybrid approaches consistently achieve high predictive accuracy (R² values up to 0.99), with feature importance analyses highlighting the roles of water-to-binder ratio, fiber content, curing regime, and supplementary cementitious materials. The integration of explainable AI methods (e.g., SHAP) has improved model transparency and practical adoption. Despite these advances, challenges remain regarding data scarcity for field-scale applications and the need for robust models that generalize across diverse environmental conditions. This review concludes with recommendations for future research directions to further enhance the reliability and applicability of ML-driven predictions for UHPC bridge infrastructure.