Dual-Granularity Clustering for Medication Recommendation with Autoregressive Contextualization

With the advancement of modern artificial intelligence technologies, the integration of Artificial Intelligence (AI) and medicine is deepening, and medication recommendation systems that utilize patients’ health records to recommend medication combinations are experiencing rapid development. Existing medication recommendation systems generally focus on the connections between multiple patient visits, while neglecting the distinction of the importance of the data points within a single visit. In addition, Intra-visit temporal information is often ignored, resulting in suboptimal recommendation accuracy. To address these limitations, we propose a medication recommendation method (DualArMed) that combines Autoregressive Model and Clustering Algorithms. Our method first employs a soft clustering algorithm to process fine-grained historical visit information, followed by an attention mechanism to assign different weights to the data points within each visit. Subsequently, we input both the soft clustering results and the original data into a hierarchical clustering module, and then leverage a graph neural network to learn the integrated coarse-grained and fine-grained information. In addition, we utilize a pretrained autoregressive model to capture the contextual relationships of each data point within a single patient visit. By combining these aspects, we obtain enhanced patient representations and improve recommendation accuracy. Extensive experiments demonstrate that our model outperforms the current state-of-the-art models across multiple metrics, with an improvement in accuracy of 4.11%.