Relational regularized risk prediction of acute coronary syndrome using electronic health records

In this paper, we attempt to utilize the information that is inherent in electronic health records (EHR) to predict clinical risks of acute coronary syndrome (ACS) patients. Because EHR data are typically highly-dimensional and non-linear, we propose a novel relational regularization-based feature selection method to identify informative risk factors from EHR data, on which a sparse ACS risk prediction model can be built. Specifically, we formulate our objective function by imposing two types of correlational characteristics, i.e., feature-feature relations and sample-sample relations, along with an l2-norm regularization term, to extract significant risk factors from EHR data. With the dimension-reduced EHR data, we train a Softmax Regression model to predict clinical risks of ACS patients. To validate the effectiveness of the proposed method, a case study was conducted on a real ACS clinical data-set that was collected from a Chinese hospital. The experimental results demonstrate the efficacy of the proposed method for improving the performance of ACS risk prediction via relational regularized risk factor selection by a comparison with state-of-the-art methods.