Heart failure disease prediction and stratification with temporal electronic health records data using patient representation

Abstract

Accurate early prediction of heart failure and identification of heart failure sub-phenotypes can enable in-time interventions and treatments, assist with policy decisions, and lead to a better understanding of disease pathophysiology in groups of patients. However, decision making more challenging for clinicians since the available data is complex, heterogeneous, temporal, and different in granularity. Even with much data, it is difficult for a cardiologist to pre-judge a patient’s heart condition at the next visit by relying on data from only one visit. Moreover, complicated and overloaded information bewilders clinicians, bringing obstacles to the stratification of patients and the mining of disease typical patterns in subgroups. To overcome these issues, this study proposes a novel Patient Representation model based on a temporal Bidirectional neural network with an Attention mechanism deep learning model called tBNA-PR. tBNA-PR effectively models heterogeneous and temporal Electronic Health Records (tEHRs) data from past and future directions to obtain informative patient representation to realize accurate heart failure prediction and reasonable patient stratification. Additionally, this study extracts typical diagnosis and prescriptions for disease patterns exploration and identifies significant features of sub-phenotypes for subgroup explanation in the context of complex clinical settings to provide better quality healthcare services and clinical decision support. This study leverages a real-world dataset MIMIC-III database. We carried out experiments on the prediction of heart failure to investigate tBNA-PR, which obtains prediction accuracy of 0.78, F1-Score of 0.7671, and AUC of 0.7198, showing a certain superiority compared with several state-of-the-art benchmarks. Moreover, we identified three distinct sub-phenotypes in all heart failure patients in the dataset with the clustering method and subgroup analysis. Sub-phenotype I has characteristics of more long-term anticoagulants. This sub-group has more patients who have the thrombotic disease. Sub-phenotype II has features of more patients having kidney disease, pneumonia, urinary tract infection, and coronary heart disease surgery history. Sub-phenotype III has characteristics of more patients having acidosis, depressive disorder, esophageal reflux, obstructive sleep apnea, and acquired hypothyroidism. Statistical tests show that the features, including age, creatinine, hemoglobin, urea nitrogen, and blood potassium, are significantly different among the three sub-phenotypes and have particular high importance. The resultant findings from this work have practical implications for clinical decision support.