The data sparsity problem caused by information overload restricts the recommendation performance of the matrix factorization model based on clinical rating data. The recommendation model integrating the clinical suggestions can effectively alleviate the sparsity of the rating data. When current recommender systems use clinical data to model patients, diseases and ADR, they only use patient clinical queries/comments on diseases and ADR as a data source. In contrast, ignoring the impact of time information on patients, diseases and ADR attributes. Aiming at this problem, a recommendation method is proposed to integrate clinical comments-level attention and temporal information (RHATR), which can fully mine the latent semantic information of clinical queries/ comments and model the dynamic changes in patients’ diseases and ADR features. By applying word-level attention to a single comment text, helpful mining information such as sentiment words and keywords in a single comment text, learning about patients’ diseases and ADR representations and applying comment-level attention to patient comment sets and disease sets with time factors, respectively. Extract practical clinical suggestions and learn dynamic representations of patients’ diseases and ADR features. The patients’ illnesses and ADR representations were known from the clinical comments and ID-based clinical details. Patient embeddings are used as final features to capture the latent factors of each patient’s diseases and ADR. Experimental results show that the proposed method achieves better results than the current baseline methods in root mean square error (RMSE) on Medhelp, Medline, Diego Laboratory and Daily Strength datasets.