Research On Disease Prediction Based On Electronic Health Record Time Series Data

Electronic health record collect data from the entire process of the patients’ visits,which is the basis for data driven disease prediction.By analyzing and mining medical time series,as an important part of Electronic health record,to discover potential disease information of patients and assist physicians in disease diagnosis and treatment of patients.In this thesis,we focus on the problem of disease prediction for medical time series.We design a feature level time aware LSTM network(FT-LSTM)to solve the sparsity and irregularity problems in medical time series,which highly improves the accuracy of disease prediction.For the problem that it is difficult to effectively combine static data features with medical time series features,we propose a multimodal features based disease prediction model(MUDIP)to successfully capture the health status of patients.In addition,we propose a class labeled time series based disease prediction model(LBDIP)to exploit the local pattern similarity feature existing in medical time series,which efficiently discovers temporal subsequences with classification capability.Main contributions of this thesis are as follows:1.Designed FT-LSTM Network:Electronic health record contain medical time series with sparsity and irregularity.Different medical variables have different collection frequencies,which leads to missing records for some of the variables,and the time intervals between consecutive variables are not always the same.However,most existing studies still consider aggregating data or filling in missing values to deal with the sparsity and irregularity problems in medical time series data,ignoring the patient health status implied by the missing information.Therefore,we design a feature level time aware LSTM network(FT-LSTM).FT-LSTM can learn the missing information of each medical variable and capture the value change and detection frequency change of medical variables,which effectively improves the accuracy of model disease prediction.2.Proposed MUDIP model:Electronic health record also includes static data,such as patient demographic data and disease data.Combining static data features with medical time series features helps to improve the predictive capability of the model.In this thesis,we propose a multimodal features based disease prediction model(MUDIP).MUDIP can extract the features of each modality data separately for medical time series data features and static data features,capture the dependency relationship between each modality feature,and effectively combine static data features with medical time series data features to comprehensively capture the patient’s health status.3.Proposed LBDIP model: Electronic health record contain medical time series with local pattern similarities.For example,doctors can determine the type of disease of a patient by finding irregular waveforms in ECG time series.We adopt a Shapelet based approach for disease prediction from the perspective of finding subsequences with classification ability in time series data.However,most existing studies have difficulty in finding high quality shapelets efficiently.in this thesis,we propose a class labeled time series based disease prediction model(LBDIP).LBDIP can efficiently find and display subsequences with classification ability in each class,which makes the model with strong disease prediction ability and interpretability.The experimental results show that the disease prediction model proposed in this thesis has better classification performance than other models.Therefore,our proposed method can help to assist doctors to diagnose diseases of patients,which is important to promote the development of medical standards in the new era.