Learning Fine-Grained Patient Similarity with Dynamic Bayesian Network Embedded RNNs

The adoption of Electronic Health Records (EHRs) enables comprehensive analysis for robust clinical decision-making in the rapidly changing environment. Therefore, using historical and similar patient records, we investigate how to utilize EHRs to provide effective and timely treatments and diagnoses for them under the circumstances that our patients are likely to respond to the therapy. In this paper, We propose a novel framework that embeds the Markov decision process into the multivariate time series analysis to research the meaningful distance among patients in Intensive Care Units (ICU). Specifically, we develop a novel deep learning model TDBNN that employs Triplet architecture, Dynamic Bayesian Network (DBN), and Recurrent Neural Network (RNN). Causal correlations among medical events are firstly obtained by the conditional dependencies in DBN, and to transmit this kind of correlations over time as temporal features, conditional dependencies in DBN are used to construct extra connections among RNN units. With specially-designed connections, the RNN is further utilized as fundamental components of the Triplet architecture to study the fine-grained similarities among patients. The proposed method has been applied to a real-world ICU dataset MIMIC-III. The experimental results between our approach and several existing baselines demonstrate that the proposed approach outperforms those methods and provides a promising direction for the research on clinical decision support.