The field of the Internet of Things (IoT) is expanding rapidly and shows the potential to completely transform the healthcare sector. The integration of machine learning algorithms in Internet of Medical Things (IoMT) systems has shown the potential to revolutionize healthcare by enabling efficient, accurate, and privacy-preserving services. However, the limitations of existing IoMT models pose significant challenges that need to be addressed to achieve sustainability in their adoption. Some of these challenges include energy consumption, complex data scheduling, low-latency models, privacy, joint offloading, limited data, and data security. There is a need to explore various novel architectures, develop new training and optimization techniques such as generative adversarial networks and stochastic gradient descent variants, and developed interpretative models to help in the decision-making process. Additionally, designing IoMT models that are suitable for low-resource environments and addressing ethical and fairness issues are crucial to preventing bias and discrimination, particularly in healthcare domains. In general, this paper reviews the current state of the art of IoMT and existing models and suggests various areas for further improvement to address the limitations of IoMT systems and enable the development of more efficient, accurate, and privacy-preserving healthcare services.