Abstract/Summary

An Electronic Medical Record (EMR) is a professional document that contains all data generated during the treatment process. The EMR can utilize various data formats, such as numerical data, text, and images. Mining the information and knowledge hidden in the huge amount of EMR data is an essential requirement for clinical decision support, such as clinical pathway formulation and evidence-based medical research. In this paper, we propose a machine-learning-based framework to mine the hidden medication patterns in EMR text. The framework systematically integrates the Jaccard similarity evaluation, spectral clustering, the modified Latent Dirichlet Allocation and cross-matching among multiple features to find the residuals that describe additional knowledge and clusters hidden in multiple perspectives of highly complex medication patterns. These methods work together, step by step to reveal the underlying medication pattern. We evaluated the method by using real data from EMR text (patients with cirrhotic ascites) from a large hospital in China. The proposed framework outperforms other approaches for medication pattern discovery, especially for this disease with subtle medication treatment variances. The results also revealed little overlap among the discovered patterns; thus, the distinct features of each pattern are well studied through the proposed framework.