In this study, we investigated which section of a page was difficult for students to read, based on eye movement data and subjective impressions of the page's difficulty, with the aim of helping teachers revise teaching materials. It is problematic to manually model relationships between eye movements and subjective impressions of the page's difficulty. Therefore, in this study, we used a neural network to model the relationships automatically. Our method generated relevance maps representing locations where students found difficulty, in order to visualize region-wise page difficulty. To evaluate the quality of the relevance maps, we compared them with a distribution of gaze points and highlights added by the students. In addition, we administered a questionnaire to evaluate whether the relevance maps were useful to teachers when revising teaching materials. Results imply that our method can provide useful information for teachers making revisions to teaching materials.