Clinical Risk Prediction Based On Multivariate Time Series EMR Data

Accurately predicting the clinical risk based on patient’s multivariate time series Electronic Medical Records(EMRs)is essential for early intervention,timely treatment of critically ill patients and allocation of medical resources.However,the patient’s EMRs usually consist of a large amount of heterogeneous multi-variate time series data such as laboratory tests and vital signs,which are produced irregularly.In addition,there is not only temporal dependence but also multiple semantic relationship between the records of each visit of a patient,and this relationship will exist across time points.For example,if a patient is diagnosed with arrhythmia in a visit,then the disease is likely to affect the subsequent diagnosis of all cardiovascular disease in this patient.In this paper,the prediction of clinical risk of patients is concretized into the tasks of mortality risk prediction and disease risk prediction.According to the characteristics of EMR data,a mortality risk prediction model based on irregular heterogeneous time series data and a disease risk prediction model based on multi-domain graph neural network are constructed.The main research contents of this paper are as follows:(1)Build a mortality risk prediction model based on irregular heterogeneous time series data.In view of the inherent irregular sampling characteristics of EMR data and the correlation between different types of data,the model introduces a time-aware Transformer to learn the personalized irregular temporal patterns of medical events.A hierarchical attention mechanism is deployed to get the accurate patient fusion representation by comprehensively mining the interactions and correlations among multiple types of medical data.(2)Build a disease risk prediction model based on multi-domain graph neural network.Most existing studies on disease risk prediction using historical diagnostic records generally focus on the relationship between diseases within visits and the temporal dependence between visits while ignoring the semantic relationship between visits.In view of the above problems,a multi-domain graph neural network is designed,which combines attention mechanism and Time-aware LSTM to learn temporal features in historical diagnostic data.The visit affinity graph and visit sequence graph is constructed based on the Pearson coefficient and time interval between medical records,and semantic features of visit sequence data is extracted through the weighted graph neural network.An adaptive feature fusion method based on attention mechanism is introduced to realize the fusion temporal features and multiple semantic features.The two clinical risk prediction models proposed in this paper are validated on the public datasets MIMIC-III and MIMIC-IV,and show better comprehensive performance compared with other excellent models.In addition,ablation studies also demonstrate the importance and validity of the core components of each model for clinical risk prediction.