Patent documents serve as a valuable source of information for companies and researchers when making decisions. Therefore, there is a demand for classifying patent documents and analyzing them. Topic modeling is a major approach to automatically extracting topics from documents, and it can be used for document clustering. Although we can apply existing topic modeling methods to the text body of patent documents for our intended purpose, most models do not fully take into account the relationships between patent documents and metadata entities, such as titles, abstracts, and classification codes. This paper explores a topic modeling approach that considers the relationships between patent documents and metadata entities, by developing and evaluating a novel method, named Metadata-Assisted Topic Modeling (MATopic). MATopic (1) generates more coherent and diverse topics by considering both the metadata of patents and their text, and (2) infers evidence about what components make up the topics by representing the detected topics as nodes in a knowledge graph. The experiments using real-world patent document datasets show that MATopic outperforms existing methods in topic diversity and coherence.