Utilizing electronic health records to predict multi-type major adverse cardiovascular events after acute coronary syndrome

摘要

Acute coronary syndrome (ACS), as a leading cause of death worldwide, is responsible for over 1.5 million hospital admissions in China each year. Major adverse cardiovascular event (MACE) after ACS often occurs suddenly resulting in high mortality and morbidity. Timely and accurate prediction of MACE after ACS could assist clinicians to pay more attention to high-risk patients and improve the quality and efficiency of care accordingly. Traditional prediction of MACE after ACS is based on aggressive medical therapy and guideline-driven treatment of a small set of coronary risk factors and is designed to predict one specific type of MACE. With the rapid development of electronic health records (EHR), more and more data-driven approaches have been proposed to explore the huge potentials of EHR data. However, existing data-driven approaches have typically been conducted by building one common predictive model and neglecting the relations among multi-type MACEs. In this paper, we propose a novel boosted regularized multi-label learning approach for multi-type MACE prediction. In detail, we formulate the prediction problem of multi-type MACE as a multi-label classification problem and incorporate the relational information among multi-type MACEs into modeling by using regularization terms to encode the corresponding relational information. In addition and to address the problem that some types of MACE are rarely occur, we utilize a boosting weighted sampling strategy to iteratively select a small subset of patient samples to train new multi-type MACE prediction models that can correct the previously wrongly predicted patient samples. A case study was conducted on a real ACS clinical dataset consisting of 2930 patient samples and collected from a Chinese hospital to validate the effectiveness of the proposed model. The performance of our best model remains robust and reaches 0.7 and 0.640 for both ischemic and hemorrhagic event prediction, respectively, in terms of AUC, and had over 2.7% and 23.1% performance gain in comparison with the state-of-the-art GRACE and CRUSADE model, respectively. The experimental results show the efficiency of our model in improving the performances on multi-type MACE prediction after ACS, by comparing with both traditional ACS risk stratification models which are widely adopted in the medical domain and state-of-the-art multi-label learning methods without efforts on incorporating the relational information of multi-type MACE into learning. In addition, we illustrate some interesting results pertaining to predictive risk factors to specific types of MACE, some of which are not only consistent with existing medical domain knowledge, but also contain suggestive hypotheses that could be validated by further investigations in the medical domain.