The importance of transfer learning as the basis for improving the accuracy of identifying the stage of Alzheimer’s disease and treating its symptoms is widely known, especially given the progressive and subtle nature of the disease. Conventional computational methods for AD diagnosis cannot efficiently diagnose early- and intermediate-stage AD and need large amounts of labelled data. These problems are reduced by transfer learning, which utilize knowledge and models from other related fields and usable in diagnostics even while data is scarce. First and foremost, a critical analysis of previous relevant transfer learning models addressing the AD stage identification problem is conducted in this paper based on their effectiveness, robustness, and feasibility. Data diversity and domain adaptation issues are discussed and recommendations for augmenting transfer learning approaches to form accurate and highly efficient AD stage classification for varying patient populations are provided.