Self-attention and Transfer Learning-Based Clinical Recommendation System for Health Care

The typical single-domain recommendation algorithm is constrained by the weak correlation between patients and disease, the issue of patient/disease cold start, and the fact that it only models item evaluations based on patients while excluding the review content. The healthcare recommendation system for clinical queries extracts patient/disease comment data from the supporting domain to address the sparse data issue in the target domain and boost recommendation precision. In this study, the self-attention mechanism and transfer learning are combined to create the clinical recommendation system (CRS-ST) which is proposed. The target domain and auxiliary domain knowledges are both fully integrated by CRS-ST, in contrast to prior methods. In order to model the patient’s preferences, a self-attention mechanism is first introduced. Next, information to increase the recommendation accuracy of another domain is introduced. Finally, the knowledge fusion module and the score prediction module integrate the information of the two domains to predict the score. Research on the PUB-Med clinical database reveals that, in comparison to the current healthcare recommendation models, the MAE and MSE values on the three separate healthcare datasets are increased by 8.4%, 13.2%, and 19.4%, respectively, and by 6.3%, 7.8%, and 5.6% in comparison to CRS-ST, which has higher MAE and MSE values. The self-attention mechanism’s efficiency and transfer learning’s benefits in reducing data sparsity and patient cold-start issues have been demonstrated in numerous trials.