An Evidence-Based Framework For Heterogeneous Electronic Health Records: A Case Study In Mortality Prediction

Yucheng Ruan, Ling Huang, Qianyi Xu, Mengling Feng

Abstract: Electronic Health Records (EHRs), characterized by their centralization of patient comprehensive disease and history information, hold significant promise to improve healthcare quality and efficiency. However, the heterogeneous nature of EHRs potentially affects the accuracy and reliability of predictive models. Many conventional methods analyze these data without explicitly considering their heterogeneity, potentially diminishing performance. Leveraging on the concepts of multimodal information analysis and the Dempster-Shafer theory, we propose an evidence-based learning framework that utilizes multi-sourced encoders to address the heterogeneity in EHRs and combines the multi-sourced evidence using Dempster’s combination rule. Our framework significantly outperforms conventional EHR analysis methods, demonstrating higher effectiveness on two tabular encoders in mortality prediction.