Due to a lack of clinicians and therapists, patients often have to return home and attempt self-directed rehabilitation therapy. These patients tend to face challenges in identifying suitable rehabilitation methods and adhering to their prescribed schedules without professional supervision. Large Language Models (LLMs) like ChatGPT, with the ability to answer sophisticated questions in natural language, appear promising in addressing this issue by assisting rehabilitation in the absence of professionals. However, significant efforts are needed to incorporate LLMs into rehabilitation, as trustworthiness and empathy are critical to ensuring patient safety and well-being. Additionally, the interaction dynamics between LLMs and patients-potentially influencing their motivation to engage in rehabilitation exercises-is also worth exploring. In this paper, we present a comprehensive review of 19 relevant studies on LLM-based rehabilitation, and conduct detailed comparisons of their methodologies, including technical approaches, interaction dynamics, and evaluation processes. Our aim is to understand how researchers are integrating LLMs into rehabilitation applications, thereby highlighting both advancements and limitations in this emerging field.