Application of interpretable deep learning in brain age prediction: a comprehensive review

Brain age prediction has emerged as a promising approach for characterizing interindividual differences in brain aging. However, the increasing complexity of deep learning models has created a substantial gap between predictive performance and interpretability. Explainable artificial intelligence (XAI) has, therefore, become increasingly important for identifying the neuroanatomical features underlying brain age predictions. This review systematically examines interpretable deep learning approaches for brain age prediction. A systematic search of PubMed, Web of Science, and ScienceDirect identified 32 eligible studies, which were organized into post-hoc explanation methods and intrinsically interpretable models. We further propose a four-dimensional operational framework encompassing Algorithm Reliability, Robustness and Reproducibility, Biological Validation, and Clinical Utility. The reviewed literature indicates a transition from localized feature attribution toward generative, disentangled, attention-based, and anatomically meaningful representations. However, interpretation remains sensitive to methodological and imaging-related factors, while brain age gap (BAG) is also influenced by age dependency, calibration, reference-population composition, and site heterogeneity. Future research should emphasize standardized evaluation and prospective clinical validation to develop reliable and clinically meaningful brain age biomarkers.