Abstract
In the study of clinical decision-making, prediction of future clinical events of patients has become an important task, especially for variant disease predictions. In previous studies, the disease prediction problems are considered as binary classification based on the patients’ electronic health records (EHRs), which lack the capacity to predict multiple types of diseases. In this paper, we propose a method which can predict both the patients’ disease types among various candidate diseases and patients’ next hospital visit time. The next hospital visit time is crucial for medical experts in making decisions, because it reflects the onset time information of disease and provides sufficient information on the severity of the disease. Our proposed method is implemented based on the point process framework, which utilizes meta-learning to gain the prior knowledge of the individual patient’s clinical data with context information, adopts sparse-attention to determine the importance of past major clinical events, and simulates the intensity of clinical events through Hawkes process to predict the types of diseases diagnosed by the doctor and patient’ next hospital visit time. The experimental data are extracted from the public datasets: Multiparameter Intelligent Monitoring in Intensive Care (MIMIC-II) and Medical Information Mart for Intensive Care (MIMIC-III). Compared with the baseline time series models, our proposed method has achieved superior results, with a higher F1-score (66.67%) and a lower root-mean-square error (RMSE) (6.69) on the test set, which proves the effectiveness of the proposed method. We further study the self-attention mechanism based on Transformer and sparse-attention methods to demonstrate the validity of our model. Our proposed method provides empirical evidence of its ability in facilitating the decision-making process of clinicians, which can be potentially utilized as effective clinical decision support tools to better improve the quality of medical services and reduce medical errors.