DKGC-LSTM:Fusion of Domain Knowledge to Guide CNN and LSTM for Heart Failure Risk Prediction

In the medical field, integrating domain-specific knowledge into disease risk prediction has become increasingly important. Despite significant advancements in deep learning techniques for handling time-series electronic health records (EHR) data, current methods often overlook the integration of domain knowledge with patient data. This paper proposes a novel method, DKGC-LSTM, for heart failure prediction based on time-series EHRs. The proposed approach utilizes domain-knowledge-guided Convolutional Neural Networks (CNN) and Long Short-Term Memory Networks (LSTM).The method consists of two main components: First, LSTM is employed to process the detection values of heart failure patient features. To address irregular time intervals in EHR data, positional encoding is incorporated into the LSTM units. Second, associations between heart failure features and the disease are extracted from a medical knowledge graph. These associations are combined with laboratory test values to form feature-disease triples, which are then processed using a CNN model to extract high-dimensional features representing the relationships between the disease and the features.Finally, these two components are integrated to construct a cascading prediction model. Experiments on the MIMIC-IV heart failure dataset demonstrate that, compared to traditional methods, this model significantly enhances prediction accuracy and recall when handling continuous data.