Research On Disease Prediction Methods For Irregular Time-series EHR Data

In recent years,with the popularization and application of Electronic Health Records(EHR)systems,a large number of electronic health records of patients have been produced.These data consist of patients’ historical visits and generally include demographic information,diagnoses,medications,laboratory tests and results,and clinical records,etc.Deep learning techniques offer improved performance and require less time for pre-processing and feature engineering than traditional machine learning and statistical techniques.The availability of large amounts of EHR data and the growing popularity of deep learning techniques have led to an increasing number of studies using deep learning techniques to apply EHR data to clinical informatics.One of these studies is disease prediction.This study has important significance.By predicting the patient’ s disease potential in the future,it can help doctors to adjust the treatment plan and effectively prevent the occurrence of disease.The most important feature of EHR data is the irregularity of patient visits.Deep learning methods have achieved success in disease prediction using EHR data.Most of the existing methods have some limitations.On the one hand,most of the methods adopt a homogeneous decay way to deal with the effect of time interval on patient’s previous visits information,and do not fully explore the effect of irregular time intervals existing in EHR on prediction results.On the other hand,the unstructured text information in EHR contains a description of the patient’s past medical history and current symptoms of the disease,so it is necessary to consider unstructured text information in disease prediction task.In addition,current methods generally consider the entire sequence of patient visits at the global level,focusing only on the coarsegrained temporal information of the sequence and overlooking fine-grained temporal information.In order to address the above limitations,this thesis provides an intensive study on disease prediction task based in irregular temporal EHR data.1.We propose a Time Interval Uncertainty-Aware and Text-Enhanced Based Disease Prediction model(TUT).The model applies cross-attention mechanism to synthetically mine disease information and unstructured text information in EHR to enhance the characterization ability of EHR data.In addition,the model uses an attention mechanism to capture the positive,neural,or negative effects of irregular time intervals on patient’s historical visits.Finally,the model uses the effects of irregular time intervals on patient’s historical visits and patient’s historical visits sequence for disease prediction.2.We propose a Triple-View Time-Aware Contrastive Learning for Disease Prediction model(TCL).The model jointly captures the coarse-grained and fine-grained temporal dependence of patient visits based on mining the uncertainty effects of irregular time intervals in the research point 1 above.In addition,the model uses contrastive learning to explore the mutual information of patient visit sequences under the three perspectives to enhance patient representation learning and achieve optimization of the prediction model.Finally,the models proposed in this thesis are compared with various state-of-the-art models on two publicly available datasets,MIMIC-Ⅲ and MIMIC-Ⅳ,and the experimental results show that the models proposed in this paper is significantly superior.