<p>Alzheimer’s disease (AD) has become a global interest due to the increase in affected people. Researchers are looking for robust techniques to diagnose AD as well as detect it early. In this context, they rely on different medical images modalities such as Positron emission tomography (PET) and Magnetic Resonance Imaging (MRI), which have proven their efficiency in Alzheimer’s detection. Moreover, merging PET and MRI into a single machine learning (ML) framework could improve the built model’s accuracy and take advantage of the complementary information in functional information provided by PET and structural information offered by MRI. However, the process of combining PET and MRI is challenging due to their structural differences. ML techniques offer robust methods for addressing this challenge. While previous surveys have discussed ML techniques for neuroimaging data fusion, they lack a detailed comparative study and a thorough discussion of the challenges facing this field, as well as possible solutions. This paper seeks to address these gaps by providing a clear and consistent review of various state-of-the-art articles that have employed various ML techniques for multimodality fusion. Specifically, it presents an overview of MRI and PET fusion-based supervised ML techniques for AD classification. Unlike previous surveys, this study summarizes the reviewed works into a detailed comparative analysis to discuss the ML techniques utilized in these studies. Additionally, it evaluates the robustness of fusion techniques based on newly proposed metrics, including, intra-relationships (<i>IARS</i>) inter-relationships (<i>IERS</i>), same-subject-modalities-interactions (<i>SSMI</i>), complementary information (<i>CI</i>), and structural information preservation, in addition to the conventional criterion of accuracy.</p>