Additive Manufacturing (AM) has considerable potential for lightweight aerospace structures and the integrated fabrication of complex components. However, the fatigue life of AM metals exhibits pronounced scatter and uncertainty because of the coupled effects of processing parameters, microstructure, internal defects, and residual stresses, making reliable life predictions challenging. Machine Learning (ML) provides a promising approach for capturing the nonlinear relationships among processing, microstructure, defects, and fatigue life. Nevertheless, high testing costs, long data-acquisition cycles and cross-platform data heterogeneity have resulted in persistent challenges associated with small samples, high noise, and limited model generalizability. This review summarizes recent advances in small-sample ML methods for fatigue life prediction of AM metals, including physics-guided ML, mechanism-based data augmentation, multi-fidelity modeling, hybrid physics-data frameworks, transfer learning and meta-learning, uncertainty quantification, and probabilistic life prediction. Current issues related to incomplete database metadata, inconsistent evaluation benchmarks, and the disconnect between experimental design and algorithm assessment are also discussed. Finally, future directions are outlined, including standardized database development, physically consistent data fusion, and unified evaluation frameworks for engineering qualifications.