Latent Dirichlet State Predictive Clustering Model for Disease Risk Prediction in Electronic Health Records

Abstract

Electronic health records (EHRs) are a valuable source of data that helps to understand patients’ health conditions and generate healthcare decisions. However, modeling the longitudinal and temporal dependencies of EHRs is challenging in disease risk prediction (DRP). To overcome this problem, this study proposed a latent Dirichlet state predictive clustering (LDSPC) using medical notes for DRP in healthcare. This process includes three modules, such as posterior, prior, and likelihood. The posterior module utilized an attentive encoder for extracting data from unstructured medical notes. Additionally, the clustering approach is integrated into the similarity module to learn the patient’s useful representation of the latent Dirichlet state. These states are clustered into numerous cluster centers, and a weighted average is applied for risk prediction. Moreover, the MIMIC-III and N2C2-2014 datasets contain unstructured medical notes that are preprocessed by non-English characters and stop word removal processes. The LDSPC achieves better accuracy of 0.9864 and 0.9694 for MIMIC-III and N2C2-2014 datasets, correspondingly which is better when compared to knowledge-enhanced multimodal learning for disease diagnosis generation (EHR-KnowGen).