Research On Next-Period Prediction And ICU Transfer Prediction Model For Time-series EHR Data

Data mining of time-series electronic health records(EHR)using deep learning techniques is a hot topic in the field of health informatics.Sequential deep learning models based on time-series EHR data can assist physicians in performing sound clinical analyses,and physicians can use the results as a basis for making effective clinical decisions.These models can improve the overall treatment efficiency of healthcare organizations and reduce the cost of patient visits.Research on efficient mining of time-series EHR data is of great importance.However,in the task of temporal EHR data mining,existing sequence deep learning models face several challenges:temporal EHR data usually consist of various heterogeneous features(demographic features,various examination indicators,medication usage,etc.),and existing models usually cannot efficiently capture the interdependence information in these highly heterogeneous features;in temporal EHR data,the time of adjacent medical events intervals are often irregular,which leads to the inability of sequence deep learning models based on the assumption of regular intervals to efficiently mine the temporal dependencies in the temporal EHR data.The above-mentioned problems lead to the existing sequential models not being able to efficiently mine the temporal patterns in EHR data well,which limits the representation learning ability of the models and ultimately affects the actual performance of the models.In this paper,two typical clinical prediction tasks in the field of time-series EHR data mining,the patient Next-Period prescription prediction task and the patient ICU ward transfer prediction task,are used as the research background to investigate the above problems in depth.(1)In the Next-Period prescription prediction study,this paper proposes a novel Cross Attention(CA)mechanism to capture the dependency information between heterogeneous features,and three recurrent neural network models based on CA mechanism:(Dual Sequences with Cross The DSCA network model can capture and exploit the interdependencies between heterogeneous features.In addition,this paper proposes a decaying update mechanism to handle historical interdependence information and uses hyperparameters to control the degree of their influence on the model output.(2)In the ICU transfer prediction study,this paper proposes a time-interval pattern-awareness-based network model(Multi-scale Interval Pattern-Aware Network,MSIPA).The model uses an interval-awareness-based approach to extract the temporal patterns of short,medium,and long time intervals in EHR data,and uses a scaled dot product attention mechanism to query the contextual information corresponding to the three patterns.In addition,MSIPA uses a Transformer network based on a multi-headed self-attentive mechanism to extract the underlying temporal dependence information,together with the interval pattern information for ICU transfer prediction.In the above study,extensive experiments were conducted using MIMIC-III,an authoritative public dataset in the field of health informatics.The experimental results show that compared with existing sequential deep learning models,the DSCA and MSIPA network models exhibit excellent predictive performance in the corresponding tasks,respectively,thus better providing healthcare organizations with complementary treatment decisions and improving their treatment.